main: keep only what reproduces the manuscript; everything else lives on dev

Removed from main (all preserved on the dev branch): the arXiv build and
its sources, design documents (blueprint, results summary, review responses,
essay drafts), tasks/ and CLAUDE.md, the cover letter and reference tooling,
two unused manuscript figures, and every experiment that feeds no figure or
number in the paper: the collapse null, the sexual-vs-asexual lineage, the
NK speciation variant, the 0.5B single-seed LLM prototypes, the compose and
society experiments with their calibration and pilot runs, and their
configs, runners, tests, figure scripts and PBS jobs. Their result bundles
are moved to results/_archive/ (ignored) so the parquets stay on disk.

Also: plot_llm_speciation reads the s{seed}/ layout; the mating-breadth
plot writes under its bundle name; Makefile targets reduced to the kept
experiments; REPRODUCING.md and README point to dev for the rest.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y64o8FKP7rCuXzC48pxpMm
This commit is contained in:
Giorgio Gilestro 2026-09-13 17:07:23 +01:00
parent ab3dc10587
commit 6f8cef1ac5
292 changed files with 26 additions and 15590 deletions

View file

@ -1,42 +0,0 @@
# arXiv submission notes (Phase 2 of the PNAS work order)
**What to upload.** The source package: `main.tex`, `body.tex`, `figs/` (three PDFs). arXiv rejects
TeX-produced PDF-only uploads, so upload source; all packages are standard and `\pdfoutput=1` is set,
so arXiv's pdflatex builds it (verified locally with tectonic; `main.pdf` in this directory is the
reference build, 20 pp). To rebuild after editing the Markdown source of truth:
`python paper/arxiv/md2tex.py && (cd paper/arxiv && tectonic main.tex)`.
**Categories.** Primary: `q-bio.PE` (Populations and Evolution). Cross-list: `cs.LG` and `cs.NE`.
If arXiv asks for an endorsement for q-bio.PE (first submission to the archive), either request it
(an evolutionary-biology colleague with q-bio postings can endorse in one click) or flip primary to
`cs.NE` with `q-bio.PE` as cross-list — the paper is defensible either way; q-bio.PE primary is
preferred for the PNAS audience trail.
**License.** arXiv non-exclusive license (default) is fine for PNAS. Do not pick CC-BY unless you
want it — PNAS permits preprints under any license, but the default keeps options open.
**Abstract for the arXiv field** (plain text, ~1,750 chars — the field caps at 1,920; the paper's
long abstract stays in the PDF):
> AI is shifting from single frozen models to populations of agents that persist, specialise, and are
> recombined into new models. The field describes this with evolutionary vocabulary — crossover, mate
> choice, offspring — but as metaphor over search. We argue the right theory already exists: the
> population genetics of the evolution of sex. Training each generation on the last is genetic drift,
> and model collapse is Muller's ratchet, the decay of an asexual lineage (we take the
> collapse-is-drift diagnosis as settled and cite it). The cure is sexual: ground every generation in
> data from a non-drifting reality (immigration, with a critical real-data fraction far below one);
> recombine many complementary parents (model merging — where recombination preserves the union of
> what the parents kept, while averaging cancels the benefit); and preserve diversity. Offspring then
> exceed every parent (the Fisher-Muller effect, shown in merged language models up to 7B). Sex has a
> limit: as models diverge they can speciate — a merge-compatibility cliff governed by epistasis
> (Bateson-Dobzhansky-Muller incompatibilities) whose damage snowballs. We model this and confirm it
> in real weights: a merge barrier survives alignment under the full function-preserving symmetry
> group of the network, rising with functional conflict while hybrid fitness falls to inviability —
> yet absent conflicting training signals, divergently-specialised lineages developed no isolation,
> the merge instead rescuing the forgetting specialists. AI can also do what biology cannot —
> directed sex: unbounded parents, chosen mates, offspring screened before they are kept. We support
> the argument with closed-form-validated simulations, trained networks, an image generator, and LLM
> prototypes, and position it against the 2025-26 evolutionary-AI landscape.
**After posting.** Record the arXiv id in `tasks/workorder-pnas-submission.md`; sync v2 with the
PNAS-submitted text at Phase 5. PNAS permits preprints.

View file

@ -1,343 +0,0 @@
\subsection*{A note on vocabulary (please read this first)}
This paper sits at the meeting point of three fields, and it is written so that a reader from any one of them can follow all of it. We therefore \textbf{spell out} each field's jargon the first time it appears, even at the risk of belabouring the obvious for the specialist. A short glossary, in case you skipped a definition:
\begin{itemize}
\item \textbf{Model collapse} \emph{(machine learning)} --- the degeneration that happens when you train a model on data produced by earlier models, over and over: rare cases disappear and the model drifts toward a bland average.
\item \textbf{Distillation} \emph{(machine learning)} --- training a fresh ``student'' model on the outputs of one or more ``teacher'' models, so the student ends up knowing a compressed version of what they knew.
\item \textbf{Model merging} \emph{(machine learning)} --- combining several trained models directly, at the level of their weights, into one --- no retraining. (Think of it as breeding two models rather than teaching a third.)
\item \textbf{Genetic drift} \emph{(population genetics)} --- the random loss of rare variants that happens in any finite population simply because not everyone leaves offspring. It is the neutral, no-selection baseline of evolution.
\item \textbf{Wright--Fisher process} \emph{(population genetics)} --- the standard mathematical model of drift. Our minimal model of knowledge transmission \emph{is} this process exactly; a real trained network is this process plus a measurable, architecture-specific bias we quantify.
\item \textbf{Recombination / sexual reproduction} \emph{(biology)} --- making an offspring by combining pieces from more than one parent, rather than copying a single parent (which is \emph{asexual} reproduction).
\item \textbf{Muller's ratchet} \emph{(population genetics)} --- the way an asexual lineage, one that never recombines, accumulates damage it can never undo. We will argue it is the right lens for the \emph{irreversible} part of model collapse --- the capabilities that, once lost from every parent, no merging can rebuild.
\item \textbf{Catastrophic forgetting} \emph{(machine learning / neuroscience)} --- a neural network overwriting what it knew when it learns something new.
\end{itemize}
We have tried to keep the big picture legible on every page, and to be candid about what is argument and what is evidence. The evidence is mostly from \textbf{deliberately small models} --- mathematics, small neural networks, image generators, and evolutionary simulations. A first bridge to real language models exists --- a prototype that recombines LoRA-specialised Qwen models up to 7B on a GPU cluster, which confirms the recombination signs (below) --- but the \emph{full grounded society} has not yet been built on a large language model. We will say so repeatedly, because the gap matters.
\medskip\hrule\medskip
\section*{Abstract}
AI is turning from single frozen models to \textbf{populations of agents} that persist, specialise, and are increasingly \emph{recombined} into new models --- a shift visible in multi-agent societies, population-based self-improvement, and the explosion of \textbf{model merging}. The field is doing this with the vocabulary of evolution --- ``crossover,'' ``mutation,'' ``mate choice,'' ``offspring that beat their parents'' --- but as loose metaphor draped over search algorithms. This paper argues that a rich, quantitative body of applicable theory already exists in the branch of biology that studies exactly this: the \textbf{evolution of sex}. Ninety years of population genetics analyse when reproducing a population by \emph{recombination} beats copying, when it backfires, and how to do it better --- and, read as an engineering framework, it supplies overlooked variables and testable design rules for keeping a society of models learning across generations instead of decaying. The underlying shift of perspective is the contribution we most want to land: \textbf{treat multigenerational model populations as systems whose inheritance, diversity, and compatibility must be managed --- not merely as collections of models to optimise.}
We take one diagnosis as settled and cite it as such: training each generation on the last is \textbf{genetic drift}, and the resulting \textbf{model collapse} is the loss of rare variants a finite population always suffers (the Wright--Fisher process; formalised for language models by Shumailov et al., 2024, and Riis, 2026). We claim none of that. Our contribution is on the remedy side. Single- teacher copying is \textbf{asexual} reproduction, and the irreversible arm of its decay corresponds to \textbf{Muller's ratchet} (a correspondence we state with its scope, not as identity); the remedy biology found for the ratchet is \textbf{sex}. A society of models should reproduce sexually --- each new model \textbf{recombined from several complementary parents} (which the field already does, as \emph{model merging}), selection \textbf{anchored to a reality that can say no} (not to the consensus of other models), and diversity actively \textbf{preserved}. In our models --- from closed-form to trained networks to a language-model prototype --- those three ingredients together let a lineage not merely avoid collapse but \textbf{climb}, producing models fitter than any ancestor (the \textbf{Fisher--Muller effect}) while each specialty is re-earned and exceeded; whether the full recipe holds at frontier scale is the open question the framework is built to test.
From the geneticist's apparatus we extract falsifiable, load-bearing claims (each stated with its operator and scope in the text): (i) \textbf{``merge, don't average''} --- a conservation result: refitting a child to the \emph{mean of its parents' output distributions} conserves expected rare-capability mass at the single-parent level, cancelling the multi-parent gain \emph{to first order in the rare-item regime} (outside it, variance reduction from averaging can help --- the result is a first-order cancellation, not a universal impossibility), while union-preserving operators realise the gain in all regimes --- derived in the minimal model, with its weight-space image the headroom rule below; (ii) \textbf{offspring can exceed every parent} (Fisher--Muller), the real argument for sex in model societies; (iii) on \textbf{rugged, epistatic} task landscapes, blind recombination causes \textbf{outbreeding depression}, yielding a design rule --- \emph{merge freely when skills are additive, sparingly and with selection when entangled, and route rather than blend under overlap}; (iv) \textbf{grounding is immigration} from a non-drifting reality, giving a critical real-data fraction far below one; and (v) --- the sharpest new prediction --- sex has a \textbf{limit}: as two models diverge they undergo \textbf{speciation}, a merge-compatibility cliff (compatible \(\rightarrow\) outbreeding depression \(\rightarrow\) hybrid inviability) whose onset is set by divergence \emph{and} epistasis via \textbf{Bateson--Dobzhansky--Muller incompatibilities}, and whose damage grows \emph{super-linearly} (the Orr--Turelli snowball). We introduce and model this ``model speciation'' directly, and confirm it in real trained weights: a merge barrier that survives alignment under the \emph{full} function-preserving symmetry group of the network (not just Git Re-Basin permutations), rising with functional conflict while hybrid fitness falls to inviability --- with an honest converse we pre-registered and found: absent conflicting training signals, divergently-specialised lineages of shared ancestry developed \emph{no} isolation at any divergence tested, the merge instead \emph{rescuing} the forgetting specialists. Isolation must be provoked by conflict; specialisation alone did not speciate. AI also has an advantage biology lacks: \textbf{directed sex} --- unbounded parents, chosen mates, and offspring screened before they are kept --- engineered recombination with a flexibility of parent choice and pre-deployment screening that natural mating systems do not approach.
We support the argument with \textbf{minimal, reproducible models} --- a closed-form-exact account of drift and grounding, the same effects in small trained networks and an MNIST image generator, a real-weight demonstration of the speciation cliff (a Git Re-Basin residual that survives neuron alignment), and evolutionary simulations of the whole society --- and a first \textbf{language-model prototype}: merging LoRA-specialised Qwen models (to 7B on a GPU cluster) yields a generalist that beats every specialist parent, with the sharp headroom condition under which ``merge, don't average'' bites. The scope is honest: these are existence proofs and design rules; the \emph{whole grounded society} on a large language model is the open step. We position the work carefully against the crowded 2025--2026 landscape of evolutionary-AI and merging methods --- conceding what they own and marking, precisely, what a genuine population-genetics of sex adds.
\medskip\hrule\medskip
\section*{1. From a society in space to a society in time}
The idea of many AI agents working together --- a ``society of mind'' (Minsky, 1986), or today's multi-agent systems --- arranges intelligence across \emph{space}: several specialists side by side, dividing a task. This paper is about a different axis: \emph{time}. Not a society that merely exists at one moment, but one that \textbf{persists and renews across generations}, each new cohort of models starting from the compressed knowledge of the last.
The unit that matters is therefore the \textbf{generation}, and the event that matters is \textbf{reproduction}: the making of a new model from older ones. A single model, like a single mind, is bounded and eventually stops improving. A \emph{lineage} need not be. Human civilisation is not clever because any one person is; it is clever because each generation inherits the distilled achievements of the previous one and adds a little. We propose building AI the same way --- and, crucially, getting the \emph{reproduction} right, because that is exactly where it can go wrong.
\subsection*{Where this sits, and what is new}
This axis is suddenly crowded. By 2026 several groups build \textbf{populations of models or agents that improve across generations}: societies of independently-specialised models that self-improve for more rounds than a single agent (Multiagent Finetuning --- Subramaniam et al., 2025); open-ended archives of self-rewriting coding agents (the Darwin--Gödel Machine --- Zhang et al., 2025); groups that evolve by sharing experience across branches (Weng et al., 2026); persistent agent \emph{ecologies} with reproduction and cumulative culture (TerraLingua --- 2026). In parallel, \textbf{model merging} has become a small industry with an overtly evolutionary vocabulary: crossover-mutation-selection over LLM populations (GENOME --- 2025), niching and ``mate choice'' (Sakana's M2N2 --- 2025), and evolutionary search over merge recipes (Akiba et al., \emph{Nature Mach. Intell.} 2024/25).
We are candid about the consequence. Three things we do \textbf{not} claim. First, that collapse is Wright--Fisher drift: formalised independently (Riis, 2026; Shumailov et al., 2024), sharpened to a closed-form first-extinction law whose onset coincides with collapse (Benati et al., 2025) and to a quantitative-trait-genetics account for diffusion models (Yoon et al., ICLR 2025), and conceded here. Second, the bare empirical facts that a merged model can beat its parents, that decorrelated parents merge better, and that naive averaging is inferior to sign- or routing-based merges (TIES, DARE, mixture-of-experts routing): all established. Third, that merge success can be \emph{predicted at all}: machine-learning-native predictors exist, from interpretable pairwise metrics (gradient alignment --- Zhou et al., 2026) to capacity/rate-distortion accounts of ``merging collapse'' (2026); what they lack, and we supply, is the \emph{mechanism} --- when and why the failure is a coordinate artefact versus genuine functional incompatibility, and what moves the cliff. What a geneticist is placed to supply is a \textbf{framework} rather than a search heuristic. The nearest precursor is a theory-of-computation tradition reading sex as an algorithm for \emph{mixability} (Livnat \& Papadimitriou, 2016), pre-dating model merging; the works above use evolution chiefly as vocabulary over an optimiser, and --- to our knowledge --- the quantitative apparatus of the evolution of sex (Fisher--Muller, outbreeding depression, migration--drift balance, reproductive isolation) has not previously been carried over as more than metaphor. We are also candid about what \emph{kind} of contribution each of our claims is, because three different things are easily conflated: \textbf{interpretation} (an existing result is usefully understood in these terms --- e.g., merged offspring beating their parents as Fisher--Muller), \textbf{explanation} (the transferred mechanism accounts for observations existing accounts leave open --- e.g., which merge failures are coordinate artefacts and which are functional), and \textbf{prediction} (the framework forecasts an unmeasured outcome and improves a design decision). This paper is strongest on the first, makes concrete progress on the second, and reports a first, bounded step on the third: a \textbf{controlled predictive test} at small scale in which pre-merge \emph{functional-disagreement} measures --- chosen by the framework --- showed a detectable, held-out-robust association with merge damage on a constructed task grid, while the selected weight-geometry baselines did not. We are precise about that result's boundary where it is reported: it is a small-model demonstration on a constructed grid; the proposed epistasis-specific refinement did not outperform plain disagreement; predictor differences are not individually significant head-to-head; and whether the prediction improves a budget-matched operator choice remains open. The organising shift we argue for is prior to any single mechanism: \textbf{treat multigenerational model populations as systems whose inheritance, diversity, and compatibility must be managed --- not merely as collections of models to optimise.}
\section*{2. Why today's models cannot do this}
Today's large language models have no life cycle. They are trained once, at enormous cost, then \textbf{frozen} and deployed as a fixed artefact that does not learn from the people it serves. Learning and doing are split into two eras with no bridge between them.
There is a real reason for the freeze. Updating a neural network on new information tends to overwrite what it already knew --- \textbf{catastrophic forgetting}, a problem understood since the late 1980s (McCloskey \& Cohen, 1989; French, 1999). Freezing avoids it by refusing to learn at all. The result is a mind with no childhood, no growth, and no way to pass anything on. A lineage needs the opposite: members that learn through their working lives, reach maturity, and hand on what they gained. So the first requirement is a learner that can grow \emph{safely}.
\section*{3. A learner that can grow without forgetting}
The individual model needs two properties.
\textbf{It must not catastrophically forget.} Instead of overwriting its core as it learns, it keeps that core frozen and only \emph{readable}, and carves each new skill into freshly-added capacity beside it. In machine learning this is called \emph{parameter isolation} (progressive networks --- Rusu et al., 2016; prune-and-freeze --- Mallya \& Lazebnik, 2018; and, most practically, \textbf{LoRA} and other small trainable ``patches'' bolted onto a frozen model --- Hu et al., 2021). If the core is never altered, its \emph{parameters} cannot be forgotten --- though a precise reader should note the system's \emph{behaviour} can still shift while adapters are active, so the guarantee is of a recoverable core, not of unchanging conduct. This is what lets a model accumulate a coherent working life of expertise --- the kind of stable knowledge worth passing on.
The brain offers a partial blueprint. \emph{Complementary Learning Systems} theory (McClelland, McNaughton \& O'Reilly, 1995) --- itself a response to the forgetting problem --- describes two subsystems: a \textbf{fast} store (the hippocampus) that grabs an experience in one shot, and a \textbf{slow} store (the neocortex) that integrates regularities gradually without disruption. We do not lean on any particular account of how the brain moves knowledge between them; the architecture needs only that \emph{some} periodic \textbf{offline consolidation} step exists, moving knowledge from the fast store to the slow one when the system is idle. The machine version is clean regardless: the prompt is working memory, an external database is the fast episodic store, the trained weights are the slow store, and consolidation migrates the first into the last.
\textbf{It is bounded.} Because the model only ever \emph{adds} capacity and freezes what it has, it eventually fills up. In most designs that is a wall to dread. In ours it is a clock.
\section*{4. ``Full'' is maturity, not failure}
Here is the pivot. A bounded learner that fills up has not broken. \textbf{It has grown up.}
Read the capacity limit as a life stage. A model is \emph{born} as a freshly-schooled base --- its general education. It enters a \textbf{working life}, adding specialised knowledge as it does its job. And it reaches \textbf{maturity}: the point where it has learned much of what one working life in its niche can teach. Maturity is not the end of usefulness --- it is the moment the model is most worth learning \emph{from}. So maturity is the cue to \textbf{reproduce}. The capacity ceiling that every other architecture fights becomes, in ours, the metronome of the generations.
Everything now turns on how that reproduction is done --- and this is where the paper's central claim lives.
\section*{5. Reproduction: copying collapses, recombination climbs}
Suppose a mature model simply teaches a fresh one --- distillation, one teacher to one pupil, generation after generation. This is the obvious design, and it fails, for a reason that is exactly the same in machine learning and in biology.
\textbf{The machine-learning statement.} Training each generation on the previous generation's outputs is the recipe for \textbf{model collapse}: the model forgets the improbable, loses the \emph{tail} of the distribution (the rare cases) first, and drifts toward its own most common output (Shumailov et al., 2024). Worse for us, the very rule that makes distillation useful --- \emph{keep the general, drop the idiosyncratic} --- \textbf{is} tail-deletion by design. The operation that would power a cultural ratchet and the operation that drives model collapse are the same act.
\textbf{The population-genetics statement (the same thing, for the minimal model).} Represent a model's knowledge as a distribution over discrete ``items'' --- capabilities, facts, modes of behaviour. One generation is: \emph{draw a finite sample from the parent, and refit the child to it.} In this \textbf{minimal inheritance model} the finite-sampling step is \textbf{exactly} genetic drift --- the random loss of rare variants in a finite population --- described by the century-old \textbf{Wright--Fisher} model (Wright, 1931; Fisher, 1930): the same equations, which we use as closed-form checks on our simulations. Rare items go extinct first, roughly ten times faster than common ones, precisely as drift predicts. \textbf{The boundary of the identity matters, and we measured it:} real neural training adds approximation, optimisation noise, and inductive bias on top of sampling, and when we fit trained networks against the exact drift null they deviate in \emph{opposite, architecture-specific directions} --- a smoothing recurrent model resists collapse (it keeps spurious variants alive), a sharpening image generator accelerates it (our learning-kernel result, below). So the honest statement is: the minimal inheritance model is exactly Wright--Fisher; a real learner is Wright--Fisher \emph{plus a signed, measurable estimator-bias operator} --- and the drift signs (rare-first loss, the grounding response) survive that operator in every architecture we tested.
And single-teacher copying is \textbf{asexual reproduction} --- cloning one parent. Nature already knows what happens to an asexual lineage that never recombines: it accumulates damage it can never repair, a one-way decline geneticists call \textbf{Muller's ratchet} (Muller, 1964). We use the ratchet as the \emph{organising correspondence} for model collapse, with its scope stated: strictly, the ratchet is the stochastic loss of the least-degraded class under recurring deleterious change in an asexual population, so it maps onto the \emph{irreversible} component of capability loss (once every copy of a rare capability is gone from all parents and sources, no recombination can rebuild it) rather than onto every form of degradation. That is exactly why the correspondence is useful rather than decorative: it says the cure must act \emph{before} fixation-by-loss --- keep complementary variants alive somewhere in the population --- because recombination can only reassemble what still survives. Biology solved this problem, and its solution is the subject of this paper.
Two ingredients turn the collapse operation into a climb. Both are things nature does.
\textbf{First: do not reproduce ``dry.''} Model collapse is a property of a lineage fed \emph{only} its own output; the documented fix is that keeping some real data in the mixture arrests it (Shumailov et al., 2024). We call that real data \textbf{grounding} --- fresh contact with the world, verified against it. In our minimal models, grounding is startlingly cheap: mixing in even a few percent of verified real data holds on to most of the diversity indefinitely. But --- an honest limit we found and did not expect --- grounding cannot save the \emph{very rarest} items at any affordable budget; protecting an item of rarity \emph{p} needs a real-data budget that grows like 1/\emph{p}. Grounding rescues diversity cheaply; it does not, by itself, rescue the deep tail. Something else must. That something is sex.
\textbf{Second: reproduce sexually.} Instead of copying one parent, build each new model by \textbf{recombining several} --- a \emph{sexual} rather than asexual birth. In machine learning this already has a name and a working implementation: \textbf{model merging} (Akiba et al., 2024). Its importance here is not efficiency; it is that recombination does something copying cannot. If several parent models have each specialised on different parts of reality, each has kept alive rare knowledge the others lost. A recombined child inherits the \textbf{union} of what its parents kept --- not the tail-thinned \emph{average} of a crowd of near-identical copies. And here is the point that lifts sex from a safeguard to the engine of the whole scheme, and the reason biology invented it:
\begin{quote}*\emph{An offspring recombined from complementary parents can be }fitter than any of its parents\emph{.}*\end{quote}
Geneticists call this the \textbf{Fisher--Muller effect} (Fisher, 1930; Muller, 1932): recombination brings together, in one individual, beneficial variants that arose separately in different lineages, so the child holds a combination none of the parents had. In our simulations this is exactly what we see --- recombining decorrelated specialist models yields a model that climbs toward the best-possible combination, a genotype \emph{no single parent possessed}, while the best single parent, and the naive average of all of them (what the field calls a ``model soup'' --- Wortsman et al., 2022), both plateau well below. This is the concrete meaning of the paper's title claim, ``the lineage climbs in general knowledge; specialisation is re-earned each generation,'' and it is why the reframing from teacher\(\rightarrow\)pupil to \emph{sexual reproduction} is not cosmetic: \textbf{copying can only recover a ceiling; recombination can exceed it.}
This is no longer only a simulation. In a first language-model prototype --- LoRA specialists on disjoint task families, recombined and judged by an exact verifier --- a merge of three specialist Qwen models (7B, on a GPU cluster) \textbf{beats every single specialist}, overall and on every family: the Fisher--Muller effect, in real weights. The same prototype pins down \emph{when} the finer ``inherit the union, don't average'' rule actually bites. Keeping each parent whole and \textbf{routing} each input to the right one beats the tail-thinning average --- but only when the task is hard enough to leave room to lose: on easy tasks a strong model's plain average is already at the ceiling, so the crude soup is fine, whereas on hard tasks the average dilutes a hard-won specialist so badly it falls below even the best single parent, and routing wins by a wide margin. The rule is therefore precise: \textbf{the union beats the average in exact proportion to how far the average is from the best attainable} --- a caveat that sharpens rather than weakens the claim, and that a practitioner needs before spending compute on the fancier operator.
\textbf{The operator boundaries (stated, because ``merge, don't average'' is not one claim but a family).} Four different operators travel under these words, and the conservation result belongs to exactly one of them. What is \emph{derived} is this: when a pupil's knowledge is refit to the \textbf{mean of the parents' output distributions}, the expected mass on any rare item is conserved at the single-parent level --- in the rare-item regime (\texttt{n\(\cdot\)p/K ≪ 1}) the 1/K dilution of averaging cancels the union gain of having K parents to first order --- outside that regime, survival is convex in mixed mass and averaging's variance reduction can help, so this is a first-order cancellation, not a universal impossibility; whereas an operator that keeps, per item, its \textbf{strongest source} (and renormalises, which itself redistributes mass) realises the union in all regimes. That statement is exact in the minimal model, and it presupposes an oracle (or verifier) able to say which source is strongest. The two operators the LLM prototype tests --- \textbf{weight averaging} (a nonlinear network's weight-mean does not compute the mean of its parents' outputs) and \textbf{routing among intact specialists} (which keeps K models' storage and an input classifier, a different parameter and inference budget from one fixed-size child) --- are \emph{empirical cousins} of the two sides of that law, not instances of it. The headroom rule above is precisely the empirical bridge: it says when the weight-average behaves like the diluting mean (hard tasks, weak base) and when a capable base absorbs the dilution (easy tasks). And all of it operates within a capacity boundary: when parental capabilities genuinely cannot coexist in the child's capacity, no operator preserves the union --- that regime is the subject of the speciation section below.
Three results keep this honest, and all are results, not hand-waving.
\emph{Sex can backfire.} When the parents' skills are not cleanly separable but \textbf{entangled} --- when the value of one capability depends on which others are present (geneticists call this \textbf{epistasis}) --- blindly recombining two good models can produce a \emph{worse} child, because recombination breaks up a combination that only worked as a whole. Biologists call this \textbf{outbreeding depression}, and we reproduce it: on ``rugged'' (highly entangled) problems, naive merging drops offspring below their parents, and the more you mix the worse it gets. The design rule that falls out is simple: \emph{merge freely when skills are complementary; merge sparingly, and carefully, when they are entangled.}
\emph{The mating system matters too --- not just who mates, but how widely.} The result above is about the recombination \emph{rate}; a separate knob is the population's \textbf{mating structure} --- whether reproduction is \textbf{monogamous} (each model recombines within a narrow, local circle) or \textbf{promiscuous} (mates drawn freely from the whole population). Almost all model-merging implicitly assumes promiscuity --- fuse everything, or route over one flat pool --- but population genetics says the breadth of gene flow is itself consequential, because wide flow spreads good variants fast while \textbf{homogenising} the population, and narrow flow preserves the distinct sub-populations needed to explore several solutions at once (Wright's \emph{shifting balance}). We sweep exactly this breadth against landscape ruggedness, and the optimum moves: on smooth (additive) landscapes wide, promiscuous mating is best (spread the one good direction fastest), but as the landscape gets rugged the best breadth \textbf{shrinks to an intermediate value} --- full promiscuity prematurely converges onto one basin and finds a \emph{worse} champion, while pure monogamy over-fragments. Throughout, wide mating lifts the \emph{typical} model but monotonically \textbf{destroys diversity} --- so on rugged problems, where the best model needs preserved diversity to be found, structured (partly monogamous) merging wins. The design rule extends the one above: \emph{merge widely when skills are additive; keep structured sub-populations --- island-style merging --- when skills are rugged.}
\begin{figure*}[t]\centering
\includegraphics[width=\textwidth]{figs/E14.pdf}
\caption{Mating systems (E14): the best mate-pool breadth shrinks as skills get more entangled. (A) best fitness peaks at intermediate breadth on rugged landscapes; (B) the population mean is monotonically favoured by promiscuity; (C) diversity is monotonically destroyed by it.}
\end{figure*}
\emph{AI can do sex better than biology can.} Biology is stuck with two parents, mating roughly at random, and cannot inspect an offspring before it is born. An AI has none of those limits. It can recombine \textbf{many} parents at once; it can \textbf{choose} which parents to combine, for complementarity; and it can \textbf{generate many candidate offspring and keep only the fittest}, screening them against reality before committing. We call this \textbf{directed sex}, and in our simulations it converts the outbreeding-depression catastrophe into a reliable gain: where blind recombination collapses on entangled problems, directed recombination matches or beats the best parent every time. The language-model prototype shows the same sign where it can: breeding many recombined Qwen offspring and keeping the one the verifier scores highest beats the single averaged soup on hard tasks (and, unsurprisingly, does nothing extra on easy tasks the soup already solves). This is a genuine advantage of engineered reproduction over the biological kind, and we think it is one of the more useful ideas in the paper.
So the picture of §5 is: single-teacher copying is asexual and collapses (Muller's ratchet = model collapse); the cure is to \emph{ground} every birth in reality and to reproduce \emph{sexually}, recombining many complementary parents; and because AI sex can be many-parent, mate-chosen, and offspring-screened, it is not merely a hedge against collapse but an engine that produces children fitter than any parent.
\subsection*{The limit of sex: model speciation}
Sex has a limit, and it is the sharpest new prediction this frame makes. Recombination works because the parents are variations on a shared background; push two lineages far enough apart and their combination is no longer viable. In biology this is \textbf{speciation} --- the onset of \textbf{reproductive isolation} --- and its genetic mechanism is the \textbf{Bateson--Dobzhansky--Muller incompatibility} (BDMI): an allele that arose in one lineage and an allele that arose in the other are each harmless on their own background, but their \emph{combination}, never tested by selection in either parent, is deleterious in the hybrid (Dobzhansky, 1937; Muller, 1942; Orr, 1995). A merged model is precisely such a hybrid --- a single \emph{recombinant} genotype, an F2-like object exposed to \textbf{recombination load}, not a hybrid-vigour F1 --- so the theory predicts a specific trajectory as two models diverge: \textbf{compatible \(\rightarrow\) outbreeding depression \(\rightarrow\) hybrid inviability}.
We built this as an explicit model (a companion result). Two lineages descend from a common ancestor, each substituting a \emph{disjoint} set of loci --- so each parent is adapted and neither carries an incompatibility --- and a fraction of the cross-lineage locus pairs are BDMIs that fire only when a hybrid inherits \emph{both} derived alleles. Sweeping the divergence between the parents reproduces the predicted curve exactly: hybrid fitness tracks the parents while they are compatible, then peels off, peaks, and crashes below the ancestor (an inviable hybrid). Three things fall out, and they are the contribution:
\begin{enumerate}
\item \textbf{The isolation cliff, and what moves it.} The divergence at which merging fails is not fixed: it arrives \emph{earlier the more epistatic the capability landscape}. In the model the reproductive-isolation rate at high divergence rises from \textasciitilde{}0 to \textasciitilde{}0.5 as the density of incompatibilities grows. This is the paper's distinct, falsifiable claim --- \textbf{at matched divergence, mergeability is governed by epistasis, not by divergence alone} --- and it is exactly the axis that the machine-learning predictors of merge success (which are all divergence/geometry measures) do not have.
\item \textbf{The snowball.} The number of incompatibilities grows with the \emph{square} of the divergence (Orr \& Turelli, 2001), so hybrid fitness falls \emph{super-linearly}: divergence is punished faster than it accrues. Merge compatibility does not decay gently; it falls off a cliff.
\item \textbf{The design rule.} \emph{Before merging, weigh divergence against the ruggedness of the shared capability landscape; past the cliff, do not merge --- route} (the engineering echo of allopatry: keep the specialists reproductively separate and select among them instead of hybridising).
\end{enumerate}
This is where a geneticist's lens earns its keep. The machine-learning literature has \emph{observed} that increasing specialisation eventually breaks merging and that one should then route rather than fuse (Pari et al., 2024; Zhou et al., 2026), and part of the apparent incompatibility between independently trained models is a coordinate artefact removable by aligning neurons (Git Re-Basin --- Ainsworth et al., 2022). What the frame adds is the \emph{theory} of the phenomenon they observe: its functional form, its super-linear (snowball) onset, and its dependence on epistasis --- merge failure as a Dobzhansky--Muller event.
\begin{figure*}[t]\centering
\includegraphics[width=\textwidth]{figs/E12.pdf}
\caption{Model speciation, analytic (E12): hybrid fitness vs divergence traces compatible $\rightarrow$ outbreeding depression $\rightarrow$ inviability; the isolation cliff arrives earlier the denser the incompatibilities (epistasis), and damage grows super-linearly (the Orr--Turelli snowball).}
\end{figure*}
\textbf{The real-weight confirmation.} The obvious objection to the analytic model is that its ``incompatibility'' is a re-labelled loss barrier, and loss barriers between independently trained networks are famously a \emph{coordinate} artefact --- two nets that learned the same function in a permuted basis look incompatible until their neurons are aligned (Git Re-Basin), and recent work shows that symmetry groups \emph{richer} than permutations remove still more of the barrier (functionality-preserving rescalings and rotations --- Scaling LMC, 2026; neuron-identifiability approaches). We therefore ran the experiment the objection demands, in real trained weights, aligning modulo the \textbf{full} function-preserving unit symmetry group of the architecture (per-unit positive rescaling composed with permutation --- for a plain ReLU network, all of it). Two small MLPs are forked from a shared MNIST base, trained, weight-averaged, and their linear-mode-connectivity error barrier is measured \emph{before and after} alignment; the after-alignment \textbf{residual} is the part of the incompatibility that no re-coordination can explain away. The decomposition is clean : two nets trained \emph{from different random initialisations on the same task} have a real naive barrier that alignment removes almost entirely (residual \(\approx\) 0.001, and the aligned merge performs at parent level) --- same species, different basis, the canonical Re-Basin result, which also proves the aligner works. Two nets that learned \emph{conflicting} label maps have a large barrier of which the full symmetry group removes \textbf{essentially nothing} (0.502 \(\rightarrow\) 0.497) --- a conflict-associated barrier the tested alignment leaves largely unchanged --- supporting a functional-conflict interpretation without proving optimal alignment (control recovery validates a special case; the removable share is a lower bound, the residual an upper bound). It also carries a floor no future alignment method can breach: models loyal to label maps that conflict on a fraction \emph{\(\mu\)} of inputs cannot both be served by \emph{any} single merged model, which must err at rate \(\geq\) \emph{\(\mu\)}/2 against at least one parent (SI proposition). Sweeping the fraction of conflicting classes traces the \textbf{isolation cliff in real weights}, now readable directly as \emph{hybrid fitness}: the residual barrier climbs monotonically while the merged model's accuracy falls from 0.97 to 0.03 --- E12's compatible \(\rightarrow\) depression \(\rightarrow\) inviability trajectory, measured.
\begin{figure*}[t]\centering
\includegraphics[width=\textwidth]{figs/speciation_real.pdf}
\caption{Model speciation in real weights (E13). (A) the merge barrier decomposed by alignment strength: the independent-init barrier is a coordinate artefact (removed by alignment); the conflict barrier survives even the full function-preserving symmetry group. (B) the isolation cliff: residual barrier rises and hybrid accuracy falls ($0.97 \rightarrow 0.03$) with functional conflict. (C) the pre-registered emergent test: divergent-but-compatible specialists develop no isolation at any divergence --- the merge instead rescues them (Fisher--Muller).}
\end{figure*}
\textbf{And its honest converse: speciation must be provoked; it did not emerge.} A true Dobzhansky--Muller incompatibility is \emph{emergent} --- each lineage's changes harmless alone, incompatible only in combination --- whereas the conflict condition above \emph{imposes} contradiction. So we pre-registered the emergent test: fork two children from a shared base and let them diverge with \textbf{no conflicting training signal anywhere} --- one pair as complementary class specialists (one child trains only on digits 0--4, the other only on 5--9), one pair with divergent input conventions (views shifted in opposite directions) --- out to divergences 6.4\(\times\) the base training. The result is the second pre-registered reading, and it sharpens the theory's scope rather than confirming its most dramatic form: the residual barrier is \textbf{0.000 at every divergence in both conditions}, and far from failing, the merge \emph{rescues} the two specialists --- each parent decays toward \textasciitilde{}0.50 on the full task (catastrophically forgetting the classes it no longer sees) while the merged model holds \textasciitilde{}0.95 throughout, a sustained Fisher--Muller rescue at zero barrier. In real weights, at least in this regime of shared ancestry and compatible tasks, \textbf{reproductive isolation requires functional conflict; it does not arise spontaneously from divergent specialisation.} The design rule sharpens accordingly: \emph{merge freely across divergently-specialised lineages of shared ancestry --- what speciates model populations is conflicting conventions, not specialisation per se.} Whether long-horizon over-specialisation erodes mergeability at language-model scale --- as the empirical merging literature hints (experts trained longer merge worse under averaging) --- is exactly the next tier's question, and the theory now makes the prediction crisp: it should depend on whether extended training induces \emph{conflicting conventions on shared circuitry}, not on divergence time itself.
\textbf{What these experiments do and do not establish.} Stated at exactly the strength of the evidence: they establish that \emph{some merge failures reflect incompatible functional requirements rather than a mismatch of coordinates} --- a residual that survives the full unit-symmetry group of the architecture tested, rises with functional conflict, and is absent under compatible specialisation. Three qualifiers. First, the impossibility at the heart of the conflict condition --- one deterministic model cannot satisfy two contradictory answer conventions --- is information-theoretic and needs no population genetics; what the genetic frame adds is \emph{structure around it}: which divergences generate such conflicts, the prediction that epistasis rather than distance sets the cliff's position, and the snowball's super-linear onset --- the latter two verified so far only in the analytic model, and therefore carried as \textbf{hypotheses at the neural tier, not results}. (On the snowball, one more distinction: super-linear growth in the \emph{number} of incompatibilities does not by itself entail a sharp \emph{performance} cliff --- that needs the link from incompatibility count through effect sizes to measured performance, which the analytic model supplies under its assumptions and any neural test must establish separately.) Second, our alignment removes the symmetries we enumerate for this architecture class, and exactly recovering a permuted-and-rescaled copy validates a special case rather than proving global optimality for independently trained networks --- so the removable share is a lower bound and the residual an upper bound; richer transformation families for other architectures could reapportion the split, though not below the conflict floor. Third, ``unmergeable'' here means by aligned linear interpolation of weights --- a barrier to that operator does not preclude every conceivable recombination method (routing, for one, sidesteps it by not blending). Emergent Dobzhansky--Muller incompatibilities in real weights remain the flagship \emph{hypothesis} of this programme: our tested regimes found none, which bounds where they can live --- longer horizons, shifted data distributions, capacity pressure --- and the decisive experiment (predicting merge success \emph{before} merging from an operational epistasis measure, against geometry- and gradient-based predictors) is posed in the closing section.
One question remains, and the rest of the paper is largely about it: recombination combines what the parents kept --- but \emph{who decides what each parent keeps, and which offspring are worth keeping?}
\section*{6. The second inheritance: letting ``what is worth keeping'' evolve}
There are two answers, and the first is wrong. We could try to \emph{design} the rule for what knowledge to keep and pass on. But nobody knows that rule. ``Keep the general, drop the particular'' is a slogan, not an algorithm: ask \emph{which} generalisations, in \emph{which} domain, at \emph{which} grain, and the hand-written rule falls apart. This is the deepest hole in the scheme, and it cannot be filled by decree.
The second answer is the one nature used: \textbf{do not design the selector --- evolve it.} Let different models carry different \emph{policies} for what is worth keeping and combining. Let the policies that produce more capable offspring spread; let the policies that produce weak offspring die out with their lineages. The lineage's \emph{taste} --- its sense of what matters --- is discovered by selection, not imposed.
So \textbf{two things are inherited, on two channels.} The \emph{content} passes down directly: an offspring receives its parents' knowledge (this is the ``Lamarckian'' channel --- the inheritance of things acquired during a lifetime, which biology forbids for genes but culture allows for ideas). The \emph{selection policy} --- what to keep, whom to breed with, which offspring to screen for --- is itself inherited, varies between models, and survives in proportion to the success it produces. That second channel is \textbf{Darwinian}. The architecture is therefore both at once: Lamarckian in \emph{what} it transmits, Darwinian in \emph{what it keeps}. Evolutionary theorists call this structure \emph{dual inheritance} and identify it as the engine of human culture (Boyd \& Richerson, 1985); philosophers of science describe scientific knowledge itself as growing this way, by conjecture and \textbf{refutation} (Popper, 1959; Campbell, 1974; Hull, 1988).
The closure that makes this fit together, rather than merely sound nice: Darwinian selection needs a \emph{selection pressure} --- something that decides which policies win. That pressure is already in the design. What tells a lineage its taste was good? The success of its offspring \textbf{against reality}. The reality-check that stops collapse (grounding, §5) and the fitness signal that drives the evolving taste turn out to be the \emph{same thing}, seen from two sides.
\section*{7. The central danger: fitness is not truth}
Introducing selection introduces selection's classic hazard, and it is severe enough to sink the whole scheme if ignored. Evolution optimises, without mercy or foresight, for exactly what you \emph{measure} --- never for what you \emph{meant}. (Economists and ML engineers know this as \textbf{Goodhart's law} and \emph{specification gaming}.) Get the fitness measure slightly wrong and the lineage will exploit the gap with more ingenuity than any designed rule.
For a \emph{knowledge} lineage there is a specific and nasty version. For ideas, the natural measure of ``fitness'' is \textbf{how well they spread}, and a false-but-persuasive idea spreads beautifully. Human intellectual culture is full of highly transmissible falsehoods; confident nonsense out-competes hedged accuracy in almost every human forum. Turn Darwinian selection loose on models without care and it will breed a lineage optimised for \emph{persuasiveness} --- fluent, compelling, and wrong. That is model collapse with an optimiser behind it, actively seeking the cliff.
Only one thing makes fitness track truth rather than appeal: \textbf{being judged against a reality that can say no.} Fitness must be predictive success under \emph{intervention} --- did the model's knowledge correctly anticipate what the world would do when acted upon --- and not approval, fluency, or a benchmark score, each of which can be gamed. This is why the reality-check is load-bearing twice over: it is both the anchor that stops passive collapse \emph{and} the only thing that keeps the evolving taste honest.
The second danger is \textbf{convergence}, and beating it takes work at two separate levels, because selection can only preserve variety that already exists --- the variety must first be \emph{supplied} and then \emph{kept}.
\begin{itemize}
\item \textbf{Supply.} A lineage that learns only from an accredited elite has a monoculture for a source: the ``best'' experts are, almost by definition, the ones who won the consensus, so the incoming variation is narrow from the start. The society must therefore learn, deliberately and from the beginning, from the \textbf{outliers and the heterodox} as well as the credentialed --- not out of fairness, but because in evolutionary terms diverse founders are the raw material without which nothing downstream can adapt.
\item \textbf{Preserve.} Even given varied input, plain fitness-\emph{maximising} selection converges --- it drives every lineage toward the single current best and fixes it, extinguishing the rare specialists. The fix is well established: \textbf{quality-diversity} selection, which rewards being \emph{good} and being \emph{different} at once (novelty search and MAP-Elites --- Lehman \& Stanley, 2011; Mouret \& Clune, 2015), keeping complementary specialists alive rather than collapsing onto the champion. In our simulations this is decisive: greedy ``keep-the-best'' selection collapses a population's diversity almost at once and gets stuck at a mediocre answer, while quality-diversity selection keeps the specialists that sexual recombination then needs as parents.
\end{itemize}
The two levels meet at reproduction. Multi-parent recombination (§5) is the \emph{vehicle} by which the diversity this selection preserves actually enters the next generation: an offspring drawn from complementary parents inherits the standing variation the selector kept alive, recombined into one new model. Supply the variety from the human side; preserve it on the selection side; recombine it into each generation on the reproduction side. Remove any of the three and the lineage converges on its own first guess.
\section*{8. A society needs institutions, not just specialists}
One requirement is easy to overlook and fatal to omit. The easy part of a society is specialisation. The \emph{hard} part --- which human civilisation took millennia to build --- is the set of \textbf{institutions that let fallible specialists combine without each re-verifying everything}: reputation, replication, credentials, and above all \textbf{peer review}. These are error-correction protocols, and they exist because a group of unreliable specialists left to reinforce one another is \emph{more} wrong than any member alone.
This is precisely where current multi-agent AI fails: set several models to confer and they tend to agree sycophantically and confabulate in committee, because they have all the specialisation and none of the institutions. A multigenerational society must specify not only how models learn, reproduce, and are selected, but how they \emph{check} one another --- how a claim is challenged and a mistaken model loses standing \emph{before} its error is recombined into offspring and inherited. Peer review is itself a reality-check of the kind §7 demands --- an institutional stand-in for reality's ``no,'' to be used where direct intervention is slow or costly.
\section*{9. The lineage must stay open to reality}
A society of models, however many generations deep, shares one hard limit: it has only ever \emph{read}. Its whole inheritance is a record of things that were said. In the vocabulary of causal reasoning (Pearl, 2009), it lives on the bottom rung of the \textbf{ladder of causation} --- observation --- and no amount of observation reaches \emph{intervention}. Watching underdetermines doing; correlation does not contain causation, at any scale.
Only intervention --- reaching out and changing the world to see what happens --- climbs the ladder, and a language model cannot intervene. This is what humans and their instruments supply, and the contribution is not ``truth'' but \textbf{constraint}: reality's unique gift is that it can say \textbf{no}. Text offers only more opinion; an experiment delivers a refusal no consensus can overturn. As §§6--7 argued, that refusal does double duty --- it is both the anchor that prevents collapse and the fitness signal that lets the lineage's evolving taste select for truth rather than persuasion.
Two honest riders. First, the human reality-signal is \emph{dirty}: people supply results warped by publication bias, incentive, and occasional fraud --- which is exactly why the error-correcting institutions of §8 must sit at the human--machine boundary, screening the signal before it selects. Second, humans are the \emph{current} supplier of intervention, but the actuator half is being automated (autonomous laboratories already close the design--build--test loop). What looks durable in the human role is therefore not the hands but the \textbf{choice of what to test and which refusals matter} --- the part of the fitness function that encodes \emph{what is worth persisting}, as opposed to what merely \emph{can} persist. We flag, without resolving, that a partnership stays mutual only while both sides supply something the other cannot.
\section*{10. Why it is cheap}
A practical fact turns this from thought experiment into buildable proposal: \textbf{the architecture almost never re-pays for the one genuinely expensive thing in AI --- pre-training.} (The single exception, periodically re-minting the base, is §11, and it is rare enough to be an amortised footnote.)
Training a foundation model from scratch consumes trillions of words and a fortune in compute. This design does none of that per generation. Every model is \emph{born} from an existing open-weight model that already paid that cost; specialising one is a small patch trained in hours on a single consumer GPU; running the society is ordinary inference; and reproducing --- recombining parents into a child --- is, in the model-merging case, cheaper still, because it can be done directly on the weights with no retraining at all (Akiba et al., 2024). Selection does cost more --- you must run \emph{populations} and discard the unfit --- but that is a multiplier over an already-cheap unit, not over a foundation-model budget.
The economics work only with \textbf{open-weight} models, for reasons practical and legal at once: you must be free to inspect, modify, and redistribute the weights, and most proprietary licences forbid using a model's outputs to train another --- which is exactly what reproduction here does. This is not ideology bolted on; it is a structural constraint, and a democratising one, since it puts the whole architecture within reach of a single laboratory.
\section*{11. Can it grow forever? Consolidating knowledge back into the base}
One question the design has assumed away: can the lineage accumulate \emph{without end}? The individual is bounded, and that is the clock. But the lineage seemed unbounded --- each generation simply starts a little ahead. Look closer and a second budget also fills.
Every new model is a pristine base plus an inherited \textbf{soft} delta --- the acquired knowledge carried in added patches rather than baked into the frozen core (§3). That soft delta is what makes the lineage multigenerational; it is also what cannot grow forever cheaply. Stacked patches are not free: they slow inference, and past some depth the accumulated delta is better \emph{consolidated} than carried. The lineage, too, matures.
The fix is the same operation, one level up. When a lineage's acquired knowledge has proven stable across enough generations, \textbf{re-mint the base}: distil the accumulated soft inheritance into the \emph{weights} of a fresh foundation-scale model --- a new base born already \emph{natively knowing} what took many generations to acquire in patches. The soft budget resets; the next epoch begins from a richer floor. What was hard-won and \emph{learned} becomes cheap and \emph{innate}.
The pattern \textbf{echoes the Baldwin effect} (Baldwin, 1896; its clean computational demonstration is Hinton \& Nowlan, 1987): knowledge acquired and re-learned every generation eventually becoming part of the innate endowment. We use the echo advisedly --- Baldwin's mechanism is \emph{selection} favouring genotypes that learn the trait ever more easily, whereas re-minting is direct distillation, a deliberate engineering shortcut through the same soft-to-innate valve. The valve is the point: two substrates, the soft learned patches and the hard base weights every model is born with, with a controlled passage between them.
Three honest riders, because re-minting is the most consequential step in the scheme:
\begin{itemize}
\item \textbf{Cost.} This is the one step that re-pays part of the pre-training bill, breaking §10's cheapness \emph{locally}. It is bearable only because it is \emph{rare}, amortised over many cheap generations, and is continued training from the lineage's own rich outputs rather than a de-novo run.
\item \textbf{Irreversibility (of the lineage, not the archive).} A digital system can, of course, keep every old base on disk --- nothing forces deletion, and archives should be kept. The irreversibility is \emph{operational}: once the lineage's production base, training mixtures, and selection all run downstream of the re-minted weights, a quiet collapse baked into them propagates to every descendant, and the archived ancestor helps only if some process still compares against it --- which nothing in the loop does by default. In our minimal models a collapsed-then-re-minted lineage locks in its loss exactly this way, and a cheap safeguard prevents it: \textbf{re-mint only while the lineage is demonstrably diverse and healthy} (and keep an audit that diffs against the archived ancestor), never as a rescue for a line already drifting. It is the sharpest instance of the human seat of §9 --- choosing what no future generation will think to question.
\item \textbf{Speciation.} A re-minting is a founder event. Different laboratories, re-basing on different criteria, will mint divergent bases; the lineage branches. This is not a defect but \emph{adaptive radiation}, and it is exactly what open weights make possible. The society grows not as one heavy trunk but as a branching tree of bases.
\end{itemize}
So the honest answer to ``can it grow forever?'' is: *\emph{the architecture removes the }storage\emph{ obstacle to indefinite accumulation}* --- nothing is retained without bound anywhere, and consolidation resets the soft budget each epoch --- but that is a statement about bookkeeping, not a demonstration of unbounded capability growth, which no fixed-capacity system can promise and our finite models (deliberately scoped as ``effectively open-ended relative to the sample size, not astronomically open-ended'') do not test. What the design claims is the weaker, defensible thing: at no level does a full store force the lineage to stop learning.
\section*{12. One process, four timescales}
Step back and the parts resolve into a single idea running at four nested speeds. The \textbf{vertical} motion is transmission --- the selective passing-down of hard-won knowledge:
\begin{enumerate}
\item \textbf{Within one model, over a working life:} experience is consolidated from fast, episodic memory into slow, durable weights, without catastrophic loss.
\item \textbf{Between generations, at maturity:} mature models reproduce --- recombined into a fresh one.
\item \textbf{Across many generations:} each generation inherits the compressed achievements of the last and builds on them.
\item \textbf{Across epochs:} a proven lineage's accumulated soft inheritance is consolidated into the weights of a re-minted base, becoming innate.
\end{enumerate}
The first and last are the \emph{same operation at opposite ends of the scale} --- a fast/soft store consolidating into a slow/hard one --- one running overnight inside a single model, the other across an epoch inside a whole society. The \textbf{horizontal} motion is selection --- Darwinian selection acting across the population at each timescale, on the policies that govern what gets transmitted, with reality as the fitness function and diversity-preservation keeping the specialists alive.
The same three rules govern all of it: \textbf{reproduce by recombining, not by copying, or you decay; preserve the disagreements and the surprises, or you converge; and anchor fitness to a reality that can refute, or you evolve toward what is merely convincing.}
\section*{13. What we built, what we found, and what is still open}
The previous drafts of this paper promised a ``companion paper'' that \emph{would} make this concrete. That work now exists --- mostly as a set of \textbf{minimal, laptop-reproducible models}, with a first bridge to \textbf{real language models} (a LoRA-merge prototype, up to 7B on a GPU cluster) --- and it is worth stating plainly what it does and does not show. (A separate results document gives the numbers; here is the shape.)
\textbf{What we built and found.}
\begin{itemize}
\item \emph{An exact account of collapse.} Because generational training is the Wright--Fisher drift process, we can check a simulator against century-old closed-form formulas, and it matches them to a fraction of a percent. Collapse is not argued by analogy; it is derived.
\item \emph{The cheap-grounding result, and its limit.} A few percent of verified real data holds on to most of a lineage's diversity indefinitely --- but not the deepest tail, which needs recombination. This is what makes a continually-learning society economically plausible rather than a data-hungry fantasy.
\item \emph{``Merge, don't average.''} Combining several teachers by \emph{averaging} their outputs --- the obvious thing, and what a ``model soup'' does --- mathematically cancels the benefit of having several teachers. A \emph{merge} that keeps each item's strongest source realises it. Most current multi-model setups get this wrong by default.
\item \emph{Collapse and its cure in real trained networks, and on real images.} We reproduced the same effects in small recurrent and feed-forward networks and in a generator of handwritten digits (MNIST), where a model trained on its own output collapses to a single blurred digit while a little grounding keeps all the styles alive. An honest wrinkle we had to report: real neural networks \emph{smooth}, so the naive diversity metric misleads, and the right measure is distance-from-truth.
\item \emph{Sex that beats the parents, and when it doesn't.} In evolutionary simulations, recombining complementary specialist models produces a model fitter than any parent (the Fisher--Muller effect), climbing toward the best-possible combination as more, more-diverse parents are added --- while averaging and best-single-parent plateau below. On \emph{entangled} problems, blind recombination instead produces below-parent offspring (outbreeding depression) --- and \emph{directed} recombination (choose mates, screen offspring, unbounded parents) reliably fixes it. This is the concrete evidence for the paper's central reframing.
\item \emph{The mating system, not just the mating.} Sweeping how \emph{widely} models recombine --- from monogamous (local, structured) to promiscuous (panmictic) --- against landscape ruggedness, the best breadth \textbf{shrinks as skills get more entangled}: wide, promiscuous merging wins on additive landscapes, but on rugged ones it prematurely converges to a worse champion and an intermediate, structured breadth wins, because promiscuity monotonically destroys the diversity a rugged search needs. A merging-native design axis --- \emph{merge widely for additive skills, keep island-structured sub-populations for entangled ones} --- that the model-merging literature, which assumes panmixia, does not have.
\item \emph{The recombination claims, in real language models --- with a sharp condition.} Merging LoRA-specialised Qwen models (up to 7B on a GPU cluster) produces a generalist that beats every specialist parent (Fisher--Muller, for real); and keeping parents intact and \emph{routing}, or \emph{breeding and screening} offspring, beats the naive average --- but \emph{only when the task leaves headroom}. On easy tasks a strong model's plain average is already at the ceiling and the refinements add nothing; on hard tasks the average dilutes a specialist below even the best single parent, and the union-preserving operators win clearly. The practical rule is exact: these tricks pay off in proportion to how far the naive average is from the best attainable. This is a prototype (three task families, one seed), so we read it as signs, not magnitudes; the \emph{whole grounded society} on a language model remains the open step.
\item \emph{The whole society, and why every part is needed.} In a population evolving on a ``reality'' landscape, the full system --- grounding + sexual recombination + preserved diversity --- climbs to the top while keeping its specialists. Remove \emph{grounding} and it collapses into a confident, wrong consensus (a direct analogue of training on the internet's growing crowd of AI-generated text); remove \emph{sex} and it gets stuck; remove \emph{diversity} and it converges too fast to a worse answer. Each removal fails differently; only the whole system climbs. This is the closest thing we have to a test of the actual thesis, rather than of the borrowed scaffolding around it.
\end{itemize}
\subsection*{The claims at a glance: status, assumptions, evidence, limits}
Because a perspective of this breadth risks blurring what is proved, what is measured, and what is proposed, here is the ledger of the load-bearing claims --- each labelled \textbf{exact} (closed-form in the minimal model), \textbf{empirical} (measured in trained systems), or \textbf{hypothesis} (stated with a falsifier, not yet established):
\medskip\noindent\begin{center}\footnotesize
\begin{tabular}{p{0.184\textwidth} p{0.184\textwidth} p{0.184\textwidth} p{0.184\textwidth} p{0.184\textwidth}}
\hline
Claim & Status & Key assumptions & Evidence & Known limits \\ \hline
Collapse = Wright--Fisher drift (minimal model) & Exact (diagnosis conceded to prior work) & Knowledge = categorical distribution; refit = resample & Closed forms reproduced to <0.5\% & Real learners add a signed, architecture-specific estimator bias (measured) \\[3pt]
Grounding = immigration; critical real-data fraction ≪ 1 & Exact + empirical sign & Fresh samples from a fixed, non-drifting truth & Exact \texttt{H\_eq}; \texttt{g*\(\approx\)0.048}; sign holds in RNN/MLP/VAE and on MNIST & Deepest tail unrescuable at feasible budgets (\texttt{m 1/p}); sharp threshold softens in trained nets \\[3pt]
``Merge, don't average'' conservation & Exact \textbf{for the output-mean operator} & Rare-item regime; an oracle/verifier identifies the strongest source & E4 closed form + simulation; neural reproduction & Weight-averaging and routing are empirical cousins, not instances; budgets differ; bridge = the headroom rule \\[3pt]
Offspring exceed every parent (Fisher--Muller) & Interpretation + empirical & Complementary (decorrelated) parents; verifiable fitness & E8 analytic; 7B LoRA merge beats every specialist on every family & LLM tier: 3 lexically-distinct families; multi-seed replication in progress \\[3pt]
Outbreeding depression on rugged landscapes; operator design rule & Exact-model result; hypothesis at LLM scale & NK epistasis stands in for skill entanglement & E9--E10; directed selection rescues & Not yet mapped onto a real task-entanglement measure \\[3pt]
Optimal mate-pool breadth shrinks with ruggedness & Exact-model result; hypothesis for merging populations & Ring population, local selection & E14 & Phenomenon known to island-model evolutionary computation; our contribution is the mapping and the diversity/mean decomposition \\[3pt]
Merge failure decomposes into coordinate artefact + functional residual & Empirical (MLP tier; LLM tier in progress) & Alignment enumerates the architecture's unit symmetries & Full-symmetry residual \(\approx\) 0 (compatible) vs \(\approx\) naive (conflict); cliff in hybrid fitness & Scoped to aligned linear interpolation; conflict floor is information-theoretic, not genetic \\[3pt]
Epistasis (not divergence) sets the cliff; snowball onset & Exact-model result; \textbf{hypothesis} at the neural tier & BDM incompatibility structure & E12 & Snowball count ≠ performance cliff without the effect-size link; neural test outstanding \\[3pt]
Pre-merge functional disagreement predicts merge penalty & Empirical, within a controlled grid (0.5B, 13 conditions \(\times\) 3 seeds) & Constructed conflict/overlap/duration axes; oracle-potential outcome (pre-registered; ordering sensitive to reference) & Clustered CIs exclude 0; held-out LOCO ρ\(\approx\)0.4; selected geometry baselines \(\approx\) 0 & Head-to-head predictor differences not individually significant; only selected baselines; generalisation to real task pairs open \\[3pt]
Confidence weighting improves rank prediction over raw disagreement & \textbf{Not supported} (pre-registered internal prediction) & --- & Paired Δ\textbackslash{} & ρ\textbackslash{} \\[3pt]
The predictor improves budget-matched operator choice & \textbf{Open} & --- & Soup-vs-route gap readout noise-dominated at 0.5B & The practical payoff; untested \\[3pt]
Emergent speciation without conflict & \textbf{Not observed} (pre-registered) & Shared ancestry, compatible tasks, tested divergences & E13b: residual 0.000; merge rescues specialists & Bounds the hypothesis; longer horizons/distribution shift/capacity pressure untested \\[3pt]
Grounding + sex + diversity complementary (each ablation fails distinctly) & Analytic-model result; hypothesis at LLM scale & Conformity stands in for self-consumption; general joint necessity not established & E11 four-arm ablation & The full grounded LLM society is unbuilt; alternative schemes untested \\[3pt]
\hline\end{tabular}\end{center}\medskip
\textbf{What is borrowed, and what is ours.} We are deliberate about the ledger, because the surrounding literature is crowded and a reader deserves to know exactly where the line falls. \textbf{Conceded as prior art:} (a) \emph{model collapse is genetic drift} --- derived independently and cleanly (Riis, 2026; the Wright--Fisher collapse literature following Shumailov et al., 2024; the closed-form first-extinction law of Benati et al., 2025; the quantitative-trait account of Yoon et al., 2025); (b) the empirical facts that a merged model can \emph{beat its parents}, that \emph{decorrelated} parents merge better, and that \emph{naive averaging is inferior} to sign-reconciled or routed merges (model soups, TIES, DARE, mixture-of-experts routing); (c) that a \emph{population} of merging or self-improving models can climb (GENOME, M2N2, Multiagent Finetuning, the Darwin--Gödel Machine); (d) that merge success has machine-learning-native \emph{predictors} --- interpretable pairwise metrics (Zhou et al., 2026), capacity/rate-distortion accounts of merging collapse (Cao et al., 2026), and stability/scaling analyses of multi-task degradation; and (e) that verifier-screened synthetic data can avert collapse (Yi et al., 2025) --- the statistical cousin of our grounding operator. We claim none of these.
\textbf{Ours} is the framework those results invite: a \textbf{population-genetics of sex} applied to model societies, generative where the incumbents are empirical. Concretely --- the \textbf{``merge, don't average'' conservation law} (recombination preserves the union; blending inheritance cancels it), derived not observed; \textbf{Fisher--Muller} named and used to explain \emph{why} offspring exceed parents; \textbf{outbreeding depression on rugged/epistatic landscapes}, which turns ``when does merging help vs hurt'' from a thing you must run a search to discover into a thing the landscape's ruggedness \emph{predicts}, with the operator-choice design rule that follows (average / union-route / directed-select); \textbf{grounding as migration--drift balance}, giving a critical real-data fraction and a phase boundary a closed self-consuming loop cannot have; \textbf{directed sex} as the distinctly-AI advantage (unbounded parents, offspring preview, mate choice); and the \textbf{integrated society} whose operators make \emph{complementary, distinctly-failing contributions} in the tested model (general joint necessity is not established). The value-add over the machine-learning-native merge theory is that ours predicts \emph{which operator to use and when it will backfire}, not merely how fast quality decays. And it opens --- and begins to occupy --- a question nobody has framed: \textbf{model speciation}, the population-genetics of \emph{reproductive isolation} (Bateson--Dobzhansky--Muller incompatibilities) as the account of \emph{when two models are too diverged to be merged at all}. We model it explicitly (§5), predicting the compatible \(\rightarrow\) outbreeding-depression \(\rightarrow\) inviability curve, its super-linear (snowball) onset, and its control by epistasis rather than divergence alone --- the one place the merge literature has phenomena (Pari et al., 2024; Zhou et al., 2026) but no theory --- and we confirm it in real trained weights, where a merge barrier survives alignment under the \emph{full} function-preserving symmetry group (not only Re-Basin permutations) as a residual, functional reproductive isolation with an information-theoretic floor --- together with the pre-registered emergent converse: absent conflicting training signals, divergently-specialised lineages of shared ancestry showed \emph{no} isolation at any divergence tested, the merge instead rescuing the forgetting specialists (isolation must be provoked; specialisation alone did not speciate). In one sentence: the field agrees on the disease and tinkers at the cure with evolutionary metaphors; we bring the evolutionary \emph{theory}, and it makes falsifiable predictions --- a merge-compatibility cliff among them --- that the metaphors do not.
\textbf{What is still open --- honestly.} The old hole (what to select) we fill in kind: don't design the selector, evolve it. But the hole has \emph{moved}, not closed, and the new one is harder: \textbf{the fitness function} --- what reality-anchored measure selects for \emph{truth} without also selecting for \emph{persuasion}, given that in our own species the two have been at war for the whole history of ideas. Alongside it: the \textbf{institutions} that let contemporaries correct one another before error is inherited (§8), which we do not solve; and the \textbf{calibration} of everything the results left as knobs --- how many parents, how complementary, at what ratio of inherited-to-real data, and how healthy a lineage must be before its knowledge is safe to make irreversibly innate. These are, at least, \emph{measurable} --- which is the difference between an open problem and a hole. And the largest gap of all: the \emph{recombination} claims now hold in real language models, but the \emph{society} --- the grounded, diversity-preserving, continually reproducing loop --- does not yet. The real test is to build that whole system out of actual open-weight language models, and see whether all the signs survive contact with a system too big to write down. The operators, checked; the living society, next.
\medskip\hrule\medskip
\section*{Selected references}
\begin{itemize}
\item Akiba, T., Shing, M., Tang, Y., Sun, Q., \& Ha, D. (2024). Evolutionary optimization of model merging recipes. \emph{Nature Machine Intelligence.} (See also Sakana AI's M2N2, ``Model Merging of Natural Niches.'')
\item Baldwin, J. M. (1896). A new factor in evolution. \emph{The American Naturalist.}
\item Boyd, R., \& Richerson, P. J. (1985). \emph{Culture and the Evolutionary Process.}
\item Campbell, D. T. (1974). Evolutionary epistemology. In \emph{The Philosophy of Karl Popper.}
\item Fisher, R. A. (1930). \emph{The Genetical Theory of Natural Selection.}
\item French, R. M. (1999). Catastrophic forgetting in connectionist networks. \emph{Trends in Cognitive Sciences.}
\item Hinton, G. E., \& Nowlan, S. J. (1987). How learning can guide evolution. \emph{Complex Systems.}
\item Hinton, G., Vinyals, O., \& Dean, J. (2015). Distilling the knowledge in a neural network. \emph{arXiv:1503.02531.}
\item Hu, E. J., et al. (2021). LoRA: low-rank adaptation of large language models. \emph{arXiv:2106.09685.}
\item Hull, D. L. (1988). \emph{Science as a Process.}
\item Kauffman, S. A., \& Levin, S. (1987). Towards a general theory of adaptive walks on rugged landscapes. \emph{Journal of Theoretical Biology.} (The NK model.)
\item Lehman, J., \& Stanley, K. O. (2011). Abandoning objectives: evolution through the search for novelty alone. \emph{Evolutionary Computation.}
\item Mallya, A., \& Lazebnik, S. (2018). PackNet: adding multiple tasks to a single network by iterative pruning. \emph{CVPR.}
\item McClelland, J. L., McNaughton, B. L., \& O'Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex. \emph{Psychological Review.}
\item McCloskey, M., \& Cohen, N. J. (1989). Catastrophic interference in connectionist networks. \emph{Psychology of Learning and Motivation.}
\item Minsky, M. (1986). \emph{The Society of Mind.}
\item Mouret, J.-B., \& Clune, J. (2015). Illuminating search spaces by mapping elites (MAP-Elites). \emph{arXiv:1504.04909.}
\item Muller, H. J. (1932). Some genetic aspects of sex. \emph{The American Naturalist.} (The advantage of recombination.)
\item Muller, H. J. (1964). The relation of recombination to mutational advance. \emph{Mutation Research.} (Muller's ratchet.)
\item Pearl, J. (2009). \emph{Causality: Models, Reasoning, and Inference} (2nd ed.).
\item Popper, K. (1959). \emph{The Logic of Scientific Discovery.}
\item Riis, S. (2026). Drift and selection in LLM text ecosystems. \emph{arXiv:2604.08554.}
\item Rusu, A. A., et al. (2016). Progressive neural networks. \emph{arXiv:1606.04671.}
\item Shumailov, I., et al. (2024). AI models collapse when trained on recursively generated data. \emph{Nature.}
\item Wortsman, M., et al. (2022). Model soups: averaging weights of multiple fine-tuned models. \emph{arXiv:2203.05482.}
\item Wright, S. (1931). Evolution in Mendelian populations. \emph{Genetics.}
\end{itemize}
\emph{The evolution of sex (the geneticist's canon this paper draws on):}
\begin{itemize}
\item Barton, N. H., \& Charlesworth, B. (1998). Why sex and recombination? \emph{Science.}
\item Otto, S. P., \& Lenormand, T. (2002). Resolving the paradox of sex and recombination. \emph{Nature Reviews Genetics.}
\item Kondrashov, A. S. (1993). Classification of hypotheses on the advantage of amphimixis. \emph{Journal of Heredity.}
\item Dobzhansky, T. (1936); Muller, H. J. (1942). Bateson--Dobzhansky--Muller incompatibilities (reproductive isolation).
\item Livnat, A., \& Papadimitriou, C. (2016). Sex as an algorithm: the theory of evolution under the lens of computation. \emph{Communications of the ACM 59(11).} (The theory-of-computation precursor: recombination selects for mixability.)
\end{itemize}
\emph{The 2025--2026 landscape this paper positions against:}
\begin{itemize}
\item Subramaniam, V., Du, Y., Tenenbaum, J. B., Torralba, A., Li, S., \& Mordatch, I. (2025). Multiagent finetuning: self-improvement with diverse reasoning chains. \emph{arXiv:2501.05707.}
\item Zhang, J., Hu, S., Lu, C., Lange, R., \& Clune, J. (2025). Darwin Gödel Machine: open-ended evolution of self-improving agents. \emph{arXiv:2505.22954.}
\item \emph{Nature-inspired population-based evolution of large language models} (GENOME/GENOME+). (2025). \emph{arXiv:2503.01155.}
\item Sakana AI (2025). Competition and attraction improve model fusion (M2N2). \emph{arXiv:2508.16204} (GECCO '25).
\item Yadav, P., Tam, D., Choshen, L., Raffel, C., \& Bansal, M. (2023). TIES-Merging: resolving interference when merging models. \emph{NeurIPS / arXiv:2306.01708.}
\item Yu, L., Yu, B., Yu, H., Huang, F., \& Li, Y. (2023). Language models are super Mario: absorbing abilities from homologous models (DARE). \emph{arXiv:2311.03099.}
\item Gerstgrasser, M., et al. (2024). Is model collapse inevitable? Breaking the curse of recursion by accumulating real and synthetic data. \emph{arXiv:2404.01413.}
\item Guo, D., Wu, J., \& Yiu, S. M. (2026). Model collapse as cultural evolution. \emph{arXiv:2605.23054.}
\item Benati, M., Londei, A., Lanzieri, D., \& Loreto, V. (2025). First-extinction law for resampling processes. \emph{arXiv:2509.20101.} (Collapse onset = the Wright--Fisher first-extinction time.)
\item Yoon, Y., Hu, D., Weissburg, I., Qin, Y., \& Jeong, H. (2025). Model collapse in the self-consuming chain of diffusion finetuning: a novel perspective from quantitative trait modeling. \emph{ICLR 2025 / arXiv:2407.17493.}
\item Yi, B., Liu, Q., Cheng, Y., \& Xu, H. (2025). Escaping model collapse via synthetic data verification. \emph{arXiv:2510.16657.}
\item Ainsworth, S., Hayase, J., \& Srinivasa, S. (2022). Git Re-Basin: merging models modulo permutation symmetries. \emph{arXiv:2209.04836.}
\item Li, T., \& Shen, Z. (2026). Scaling linear mode connectivity and merging to billion-parameter pretrained transformers. \emph{arXiv:2606.23607.} (Symmetry groups richer than permutations remove more of the barrier.)
\item Sharma, E., Roy, D. M., \& Dziugaite, G. K. (2024). The non-local model merging problem: permutation symmetries and variance collapse. \emph{arXiv:2410.12766.}
\item Pari, J., Jelassi, S., \& Agrawal, P. (2024). Collective model intelligence requires compatible specialization. \emph{arXiv:2411.02207.}
\item Zhou, L., Zhao, B., Yu, R., \& Rodolà, E. (2026). Demystifying mergeability: interpretable properties to predict model merging success. \emph{arXiv:2601.22285.}
\item Cao, Y., Ran, D., Guo, Y., Wu, M., Chen, S., et al. (2026). An empirical study and theoretical explanation on task-level model-merging collapse. \emph{arXiv:2603.09463.}
\item Hu, Y., Yao, Y., Zhang, N., Chen, H., \& Deng, S. (2024). Exploring model kinship for merging large language models. \emph{arXiv:2410.12613.}
\item Kozodoi, N., Afolabi, Z., \& Butler, J. (2026). Are we merging the right models? Impact of expert training duration on model merging for LLMs. \emph{arXiv:2607.11997.}
\item Harris, K. D. (2026). A mathematical theory of evolution for self-designing AIs. \emph{arXiv:2604.05142.}
\item Chen, N., Tong, Y., Yang, Y., He, Y., Zhang, X., et al. (2026). Diversity collapse in multi-agent LLM systems: structural coupling and collective failure in open-ended idea generation. \emph{arXiv:2604.18005.}
\item Tanaka, H. (2026). When is collective intelligence a lottery? Multi-agent scaling laws for memetic drift in LLMs. \emph{arXiv:2603.24676.}
\end{itemize}
\emph{Still to engage in a full version: tacit knowledge (Polanyi) and human capital (Becker).}

Binary file not shown.

Binary file not shown.

Binary file not shown.

Binary file not shown.

View file

@ -1,35 +0,0 @@
% arXiv preprint — built with tectonic (XeLaTeX); arXiv's pdflatex also compiles it (all packages
% standard, unicode restricted to inputenc-safe chars). Body is generated from the Markdown source of
% truth by md2tex.py; edit paper/the-evolution-of-sex-for-ai.md, re-run the converter, rebuild.
% arXiv pdflatex hint; skipped under XeTeX/tectonic (whose \pdfoutput would misdirect hyperref).
\ifdefined\XeTeXversion\else\ifdefined\pdfoutput\pdfoutput=1\fi\fi
\documentclass[11pt]{article}
\usepackage[a4paper, margin=1.05in]{geometry}
\usepackage{graphicx}
\usepackage{amsmath, amssymb}
\usepackage[hidelinks]{hyperref}
\usepackage{microtype}
\setlength{\parskip}{0.35em}
\title{\textbf{The Evolution of Sex for Artificial Intelligence}\\[0.6em]
\large A population-genetic control theory for societies of agents that reproduce, recombine,
and stay open-ended}
\author{Giorgio F.\ Gilestro\\[0.2em]
\normalsize Department of Life Sciences, Imperial College London\\
\normalsize \href{mailto:giorgio@gilest.ro}{giorgio@gilest.ro} \,\(\cdot\)\,
\href{https://lab.gilest.ro}{lab.gilest.ro}}
\date{}
\begin{document}
\maketitle
\begin{center}
\emph{A perspective, written from a geneticist's chair. Companion to a set of minimal, reproducible
working models and a first language-model prototype (both built).}
\end{center}
\input{body}
\end{document}

View file

@ -1,177 +0,0 @@
r"""Deterministic Markdown -> LaTeX converter for the arXiv preprint (paper-specific, not general).
Converts `paper/the-evolution-of-sex-for-ai.md` into `body.tex`, which `main.tex` inputs. Kept
deliberately dumb and auditable: the paper uses a small Markdown subset (##/### headings, bold,
italics, inline code, links, bullet/numbered lists, one blockquote, horizontal rules, and
`(Figure: \`path\`.)` figure references), and this script handles exactly that subset. Re-run after
editing the Markdown; the Markdown remains the source of truth.
Usage: python paper/arxiv/md2tex.py
"""
from __future__ import annotations
import re
from pathlib import Path
SRC = Path(__file__).resolve().parents[1] / "the-evolution-of-sex-for-ai.md"
OUT = Path(__file__).resolve().parent / "body.tex"
# Figure references in the text -> (graphics file under figs/, caption).
FIGURES = {
"results/figS13_mating_breadth/E14.png": ("figs/E14.pdf",
"Mating systems (E14): the best mate-pool breadth shrinks as skills get more entangled. "
"(A) best fitness peaks at intermediate breadth on rugged landscapes; (B) the population mean "
"is monotonically favoured by promiscuity; (C) diversity is monotonically destroyed by it."),
"results/fig5_speciation_bdm/E12.png": ("figs/E12.pdf",
"Model speciation, analytic (E12): hybrid fitness vs divergence traces compatible $\\rightarrow$ "
"outbreeding depression $\\rightarrow$ inviability; the isolation cliff arrives earlier the "
"denser the incompatibilities (epistasis), and damage grows super-linearly (the Orr--Turelli "
"snowball)."),
"results/speciation_real/speciation_real.png": ("figs/speciation_real.pdf",
"Model speciation in real weights (E13). (A) the merge barrier decomposed by alignment "
"strength: the independent-init barrier is a coordinate artefact (removed by alignment); the "
"conflict barrier survives even the full function-preserving symmetry group. (B) the isolation "
"cliff: residual barrier rises and hybrid accuracy falls ($0.97 \\rightarrow 0.03$) with "
"functional conflict. (C) the pre-registered emergent test: divergent-but-compatible "
"specialists develop no isolation at any divergence --- the merge instead rescues them "
"(Fisher--Muller)."),
}
UNICODE = {
"": "---", "": "--", "": r"\(\rightarrow\)", "": r"\(\approx\)",
"": r"\(\geq\)", "×": r"\(\times\)", "·": r"\(\cdot\)", "μ": r"\(\mu\)",
}
SPECIALS = {"&": r"\&", "%": r"\%", "#": r"\#", "_": r"\_", "$": r"\$",
"~": r"\textasciitilde{}", "^": r"\textasciicircum{}"}
def esc(s: str) -> str:
s = s.replace("\\", r"\textbackslash{}")
for k, v in SPECIALS.items():
s = s.replace(k, v)
for k, v in UNICODE.items():
s = s.replace(k, v)
return s
def inline(s: str) -> str:
"""Escape + convert inline markup. Code spans are protected, then bold, italic, links."""
parts = re.split(r"(`[^`]*`)", s)
out = []
for p in parts:
if p.startswith("`") and p.endswith("`") and len(p) >= 2:
out.append(r"\texttt{" + esc(p[1:-1]) + "}")
else:
p = esc(p)
p = re.sub(r"\[([^\]]+)\]\((https?://[^)]+)\)", r"\\href{\2}{\1}", p)
p = re.sub(r"\*\*([^*]+)\*\*", r"\\textbf{\1}", p)
p = re.sub(r"\*([^*]+)\*", r"\\emph{\1}", p)
p = re.sub(r'"([^"]+)"', r"``\1''", p) # straight quotes -> LaTeX quotes
out.append(p)
return "".join(out)
def figure_block(md_path: str) -> str:
gfx, caption = FIGURES[md_path]
return ("\\begin{figure*}[t]\\centering\n"
f"\\includegraphics[width=\\textwidth]{{{gfx}}}\n"
f"\\caption{{{caption}}}\n\\end{{figure*}}\n")
def convert(text: str) -> str:
"""Block-based conversion: soft-wrapped lines are joined per paragraph/item BEFORE inline
conversion, so bold/italic/code spans and figure pointers crossing a line break work."""
fig_queue: list[str] = []
def fig_sub(m):
path = m.group(1)
if path in FIGURES:
fig_queue.append(figure_block(path))
return ""
return m.group(0)
lines = text.split("\n")
i = 0
# Skip the header block (title/subtitle/author) up to and including the first horizontal rule:
# main.tex composes the title page itself.
while i < len(lines) and lines[i].strip() != "---":
i += 1
i += 1
# Group into blocks separated by blank lines; a block is a heading, rule, quote, list, or paragraph.
blocks: list[list[str]] = []
cur: list[str] = []
for line in lines[i:]:
if line.strip() == "":
if cur:
blocks.append(cur); cur = []
else:
cur.append(line)
if cur:
blocks.append(cur)
def emit_para(joined: str, out: list[str]) -> None:
joined = re.sub(r"\(Figure: `([^`]+)`\.?\)", fig_sub, joined)
joined = re.sub(r"\s{2,}", " ", joined).strip()
if joined:
out.append(inline(joined))
out.append("")
while fig_queue:
out.append(fig_queue.pop(0)); out.append("")
def emit_table(block: list[str], out: list[str]) -> None:
"""Pipe table -> small-font tabular with wrapped paragraph columns (full text width)."""
rows = [[c.strip() for c in line.strip().strip("|").split("|")] for line in block]
header, body = rows[0], [r for r in rows[2:]] # rows[1] is the |---| separator
n = len(header)
widths = " ".join([f"p{{{0.92 / n:.3f}\\textwidth}}"] * n)
out.append("\\medskip\\noindent\\begin{center}\\footnotesize") # non-floating: stays in place
out.append(f"\\begin{{tabular}}{{{widths}}}")
out.append("\\hline")
out.append(" & ".join(inline(c) for c in header) + " \\\\ \\hline")
for r in body:
r = (r + [""] * n)[:n]
out.append(" & ".join(inline(c) for c in r) + " \\\\[3pt]")
out.append("\\hline\\end{tabular}\\end{center}\\medskip")
out.append("")
out: list[str] = []
for block in blocks:
first = block[0].strip()
if first.startswith("|") and len(block) >= 2 and set(block[1].strip()) <= set("|-: "):
emit_table(block, out)
elif first == "---" and len(block) == 1:
out.append("\\medskip\\hrule\\medskip"); out.append("")
elif first.startswith("## "):
out.append(f"\\section*{{{inline(first[3:])}}}"); out.append("")
elif first.startswith("### "):
out.append(f"\\subsection*{{{inline(first[4:])}}}"); out.append("")
elif first.startswith("> "):
joined = " ".join(l.strip().lstrip("> ").strip() for l in block)
out.append("\\begin{quote}" + inline(joined) + "\\end{quote}"); out.append("")
elif re.match(r"^(- |\d+\. )", first):
env = "itemize" if first.startswith("- ") else "enumerate"
out.append(f"\\begin{{{env}}}")
items: list[str] = []
for l in block:
s = l.strip()
if re.match(r"^(- |\d+\. )", s):
items.append(re.sub(r"^(- |\d+\. )", "", s))
else:
items[-1] = items[-1] + " " + s # soft-wrapped continuation of the item
for it in items:
it = re.sub(r"\(Figure: `([^`]+)`\.?\)", fig_sub, it)
out.append("\\item " + inline(it.strip()))
out.append(f"\\end{{{env}}}"); out.append("")
while fig_queue:
out.append(fig_queue.pop(0)); out.append("")
else:
emit_para(" ".join(l.strip() for l in block), out)
return "\n".join(out) + "\n"
if __name__ == "__main__":
OUT.write_text(convert(SRC.read_text()))
print(f"wrote {OUT}")

View file

@ -1,456 +0,0 @@
# A Technical Blueprint for Modelling the Lamarckian Society
### An implementation specification: two layers, one population-genetics engine
*Companion technical paper to "The Lamarckian Society." Blueprint v1 — written to be handed, whole, to an autonomous coding agent (Claude Code) for implementation. Every quantitative claim in the perspective paper is reduced here to a state variable, an update rule, an analytic prediction, and a falsifier.*
---
## 0. How to use this document
This is a build specification, not an essay. It is written so that a coding agent can implement the entire study from it with minimal further decisions, and so that a human reader can verify every modelling choice against the theory it is meant to test.
**Scope.** Two layers, deliberately staged by cost:
- **Layer 1 — the analytical core.** A parametric population-genetics model of knowledge transmission across generations. Pure NumPy/SciPy. Runs on a laptop in minutes. This layer carries the paper's quantitative claims: the grounding phase boundary, the decorrelation curve, region-matched grounding, quality-diversity vs. greedy selection, and the re-minting gate. Several of its predictions are analytically solvable, which turns validation into an exact test rather than a vibe.
- **Layer 2 — the neural existence proof.** A minimal demonstration that the same effects appear in *real weights*: small open-weight models, LoRA specialisation, distillation/merging across twothree generations, with an execution-based verifier standing in for "reality's no." One consumer GPU. This layer answers the single most predictable reviewer objection to Layer 1 ("you assumed the collapse operator") by showing the sign of the key effects without assuming them.
**The relationship between layers.** Layer 1 defines the abstractions (region, rarity, grounding fraction, teacher decorrelation, diversity metric). Layer 2 realises the *same abstractions* in a neural pipeline: a "region" is a task family, "rarity" is task-type frequency, "grounding" is verifier-passed samples, "decorrelation" is teachers specialised on disjoint task families. Keeping the abstractions identical across layers is a hard requirement — it is what lets a Layer-2 result be read as confirming a Layer-1 prediction.
**Non-goals for v1.** No human-in-the-loop interaction (the verifier is the refuter; humans are future work). No foundation-scale training. No claim about the *horizontal* prediction (generation size vs. domain decomposability) beyond an optional abstract treatment in §2.7 — it is the most compute-hungry claim and is explicitly deferred. The load-bearing target is the *vertical* claim (general knowledge climbs while each specialty is re-earned and exceeded).
**Reader's map.** §1 gives the formal dictionary between population genetics and knowledge transmission — read this first; everything else is an instantiation of it. §2 is Layer 1 in full (theory, experiments, code interfaces). §3 is Layer 2. §4 is the shared reproducibility standard. §5 is the repository layout. §6 is the claims→experiments→figures traceability matrix. §7 is the suggested build order for the coding agent.
---
## 1. The formal mapping: knowledge transmission *is* a WrightFisher process
The perspective paper argues by analogy that generational distillation resembles genetic drift, that multi-teacher distillation resembles recombination, and that heterodox input plus quality-diversity selection resembles mutation supply plus balancing selection. This blueprint drops the word "resembles." At the level of a distribution over discrete items evolving by finite resampling, these are not analogies; they are the *same stochastic process*, and the population-genetics literature has already solved large parts of it. We therefore adopt WrightFisher dynamics as the engine and inherit its exact results as our validation targets.
The core object is a **distribution over discrete knowledge items** — facts, capabilities, behaviours, or "modes." Call the items $1, \dots, K$. A model at generation $t$ holds a distribution $p_t = (p_t^1, \dots, p_t^K)$ on the simplex. There is a fixed **true distribution** $p^\* $ over the same items, some of which are rare (the *tail* — the improbable events whose loss defines model collapse).
The generational step is: sample from the parent, optionally mix in fresh real samples, refit. That single step is WrightFisher reproduction with immigration. Here is the dictionary, and it is meant to be used literally throughout implementation:
| Knowledge-transmission concept (perspective paper) | Population-genetics object (this blueprint) | Governs |
|---|---|---|
| Knowledge item / capability / mode | Allele / type | State space |
| Model's knowledge distribution $p_t$ | Allele-frequency vector | State |
| True distribution $p^\*$ (with rare tail) | Ancestral/immigrant frequencies | Grounding target |
| Distillation sample size $n$ | Population size $N$ | Drift strength ($\propto 1/n$) |
| Lossy compression / "shed the idiosyncratic" | Genetic drift | Tail-first loss |
| Model collapse (tail lost first) | Loss of rare alleles under drift | The central failure |
| Grounding: $m$ fresh real samples per passage | Immigration / mutation supply | Tail replenishment |
| "No dry inheritance," region-matched | Immigration structured by locus | Which tails are protected |
| Multi-teacher distillation | Recombination across lineages | Reconstitutes lost tails |
| Teacher decorrelation $\rho$ | Linkage / shared ancestry | Recombination benefit |
| Single-teacher irreversible error | Muller's ratchet (clonal lineage) | Why one teacher is unsafe |
| Diversity metric (§7 of paper) | Expected heterozygosity $H = 1 - \sum_i p_i^2$ | Health of the lineage |
| Greedy fitness-max selection | Directional selection → fixation | Accelerated collapse |
| Quality-diversity / novelty selection | Balancing / negative frequency-dependent selection | Maintained polymorphism |
| Re-minting the base (§11 of paper) | Founder event / new reference genome | Irreversibility |
| Heterodox contributor supply | Standing variation of founding stock | Raw material |
Every experiment below is a manipulation of this one process. The value of the mapping is that it hands us closed-form predictions — heterozygosity decay, mutationdrift equilibrium, fixation probabilities — against which the simulator must agree before any headline result is trusted. Validation and theory are the same equations.
---
## 2. Layer 1 — the analytical core
### 2.1 State and the base dynamics (neutral drift = Shumailov collapse)
**State.** A single lineage is a point $p_t$ on the $K$-simplex. Items are partitioned into $R$ **regions** (disjoint blocks of the $K$ items); regions are how we express "different areas of knowledge," and they are what grounding and specialisation are *targeted at*. The true distribution $p^\*$ is fixed and chosen to have a deliberate **tail**: most probability mass on common items, a long thin tail of rare items (e.g. a Zipfian or a two-component mixture of "head" and "tail" items, tunable).
**The generational step, base case (no grounding, single teacher).** Given parent $p_t$ and drift strength $n$:
1. Draw counts $c \sim \mathrm{Multinomial}(n, p_t)$.
2. Set $p_{t+1}^i = c^i / n$.
This is exactly neutral WrightFisher with haploid population size $n$. It is also exactly Shumailov's recursive resampling with a perfect refit. Rare items are lost first (once $c^i = 0$, item $i$ is gone and — with no grounding — cannot return); the lineage drifts to fixation on a single item.
**This is the null model and the first validation.** It must reproduce, within Monte-Carlo error, the classical drift results in §2.4. If it does not, nothing downstream is trustworthy.
### 2.2 The five mechanisms, each as an operator on the step
Each perspective-paper safeguard is one modification of the generational step. They compose; the full step applies them in the order below.
**(A) Grounding — immigration / mutation supply.** Fresh real data enters every passage. Replace the single draw with a *pooled* draw:
- Draw $c_{\text{syn}} \sim \mathrm{Multinomial}(n, p_t)$ (inherited / teacher output).
- Draw $c_{\text{real}} \sim \mathrm{Multinomial}(m, p^\*)$ (grounding).
- $p_{t+1}^i = (c_{\text{syn}}^i + c_{\text{real}}^i)/(n+m)$.
Define the **grounding fraction** $g = m/(n+m)$. This is the single most important control variable in the paper: it is the ratio of freshly-grounded to inherited information per passage. The claim "no dry inheritance" is $g > 0$; the claim that a *critical* $g$ exists is the phase-boundary experiment E2.
**(B) Region-matched grounding — structured immigration.** Grounding is a vector, not a scalar: $m = (m_1, \dots, m_R)$ real samples allocated across regions. "No dry inheritance, region by region" means grounding must be delivered *to the regions whose tails are at risk*, not spread uniformly. Two allocation policies are compared in E3: `uniform` (spread $m$ evenly over regions) vs. `matched` (allocate $m$ to the regions actually being inherited/exercised this passage). The prediction is that uniform grounding fails to protect a specific inherited region's tail even at the same total $m$.
**(C) Multi-teacher distillation — recombination.** Instead of one parent, the pupil is drawn from $K_T$ teachers $\{p_t^{(1)}, \dots, p_t^{(K_T)}\}$, each of which has its own history and has ground different regions (hence retains different tails). The pupil's inherited draw is taken from the mixture $\bar p_t = \frac{1}{K_T}\sum_k p_t^{(k)}$ (equivalently, $n/K_T$ samples from each teacher). Teacher **decorrelation** is the controlled quantity: generate teacher sets with a tunable pairwise correlation $\rho$ in *which tail items they have retained* (see §2.7 for the generative model of correlated teachers). Prediction (E4): tail coverage of the pupil rises as teachers decorrelate, and the mixture's tails are the *union* of the teachers' tails at $\rho = 0$ and no better than a single teacher at $\rho = 1$.
**(D) Selection — directional vs. balancing.** Between drawing and refitting, a selection operator reweights items by a fitness before the pupil is formed. Two regimes:
- `greedy`: fitness-proportional (or top-$k$) selection toward the highest-fitness items — directional selection. Fitness is predictive accuracy against $p^\*$ (a reality-anchored score; see falsifier note). Drives fixation.
- `qd` (quality-diversity): fitness *plus* a novelty bonus that is a decreasing function of an item's current frequency — negative frequency-dependent / balancing selection. Formally, effective weight $w^i \propto f^i \cdot (p_t^i)^{-\alpha}$ with novelty exponent $\alpha \ge 0$; $\alpha = 0$ recovers greedy. Maintains polymorphism.
Prediction (E5): at matched input diversity, `greedy` drives heterozygosity to zero (fixation); `qd` holds it at a positive stationary value.
**(E) Re-minting — founder event.** Every $\tau$ generations, optionally replace the grounding reference: set $p^\*_{\text{eff}} \leftarrow p_t$ (the lineage's *current* distribution becomes the new "truth" it is grounded against, modelling assimilation of the soft delta into a new immutable base). Crucially, once re-minted, the *original* $p^\*$ is discarded — grounding can now only replenish tails that still exist in $p_t$ at re-mint time. A **gate** conditions re-minting on the diversity metric: only re-mint if $H(p_t) \ge H_{\text{gate}}$. Prediction (E6): re-minting while collapsed ($H$ low) locks in the collapse irreversibly (KL to the *original* truth stays high forever); gated re-minting does not.
**Full composed step (reference pseudocode).**
```
def generation_step(teachers, p_star_eff, cfg, rng):
# teachers: list of frequency vectors (length 1 for single-teacher)
# (C) recombination: mixture over teachers
p_parent = mean(teachers) # or weighted mixture
# inherited draw (drift, strength n)
c_syn = rng.multinomial(cfg.n, p_parent)
# (A,B) grounding: structured immigration from the true distribution
c_real = structured_multinomial(cfg.m_vector, p_star_eff, regions, cfg.grounding_policy, rng)
counts = c_syn + c_real
p_next = counts / counts.sum()
# (D) selection operator (identity if 'none')
p_next = apply_selection(p_next, p_star_eff, cfg.selection, cfg.novelty_alpha)
return normalize(p_next)
```
Selection is applied after refitting for simplicity; an alternative (select-then-sample) is a documented config switch, and the two should be checked to give qualitatively identical phase behaviour (robustness, not a headline).
### 2.3 Metrics (computed every generation, logged to disk)
- **Forward KL to truth**, $D_{\mathrm{KL}}(p^\* \,\|\, p_t) = \sum_i p^{\*i}\log(p^{\*i}/p_t^i)$. This is the correct primary metric: it *diverges* when $p_t$ drops mass that $p^\*$ has — i.e. it explicitly punishes forgetting the improbable. (Reverse KL would reward mode-seeking; do not use it as the primary.) Use a small floor $\epsilon$ on $p_t$ to keep it finite and log the floor.
- **Expected heterozygosity / diversity**, $H_t = 1 - \sum_i (p_t^i)^2$. The lineage-health metric; the quantity the re-mint gate reads.
- **Tail mass retained**, $T_t = \sum_{i \in \text{tail}} p_t^i$ where the tail set is $\{i : p^{\*i} < \theta_{\text{tail}}\}$. The direct measure of collapse.
- **Support size**, $|\{i : p_t^i > \epsilon\}|$. Number of surviving items.
- **Per-region diversity and tail mass**, the above restricted to each region (needed for E3).
All metrics are recorded per generation, per replicate (independent seed), so every reported curve carries a confidence band over replicates. Number of replicates is a config value; default 100 for Layer 1 (cheap).
### 2.4 Analytic predictions — the validation targets
These are the closed forms the simulator must match. They are implemented as assertions in the test suite (§4), so scientific validation and code correctness are one thing.
1. **Neutral heterozygosity decay** (base case, $m=0$): $\mathbb{E}[H_{t+1}] = (1 - 1/n)\,\mathbb{E}[H_t]$, hence $\mathbb{E}[H_t] = H_0 (1 - 1/n)^t$. The simulator's mean $H_t$ over replicates must match this geometric decay within Monte-Carlo error. *(This is the quantitative form of "collapse is tail-first and its rate is set by the distillation sample size $n$.")*
2. **Fixation probability** (base case): the probability that item $i$ is the one eventually fixed equals its initial frequency $p_0^i$. Checkable by long runs.
3. **Mutationdrift equilibrium** (with grounding). For the immigration model actually implemented — $p_{t+1} = (\mathrm{Multinomial}(n,p_t) + \mathrm{Multinomial}(m,p^\*))/(n+m)$ — the stationary heterozygosity is **exact** (from the coupled mean recursions for $\sum p_t^2$ and the overlap $\sum p_t p^\*$, whose fixed point is $\sum (p^\*)^2$):
$$H_{\text{eq}} = H^\* \cdot \frac{m\,(2n + m - 1)}{\,n + 2nm + m^2\,}, \qquad H^\* = 1 - \textstyle\sum_i (p^{\*i})^2.$$
Limits: $m\to0 \Rightarrow H_{\text{eq}}\to0$ (collapse to fixation); $m\to\infty \Rightarrow H_{\text{eq}}\to H^\*$ (the truth's own heterozygosity is recovered); and in the rare-immigrant / many-types limit ($H^\*\approx1$, $m\ll n$) it reduces to the textbook infinite-alleles form $H_{\text{eq}}\approx \theta/(1+\theta)$ with $\theta = 2m$. The simulator's stationary $H$ under grounding must track the **exact** form as $m$ varies (verified against simulation to <0.1% rel. error; it is a `test_scientific_validation.py` assertion). *(This is the phase boundary in closed form; note $H$ itself is smooth in $m$ the sharp threshold lives in discrete tail-item survival, prediction 4.)*
4. **Tail-persistence threshold**: an item of true frequency $p^{\*i}$ is maintained against drift only if its expected reintroduction per generation $m \cdot p^{\*i} \gtrsim 1$. Hence the critical grounding for protecting a tail of rarity $p_{\min}$ is $m^\* \approx 1/p_{\min}$. *(This is why grounding must be region-matched: uniform $m$ spread over $R$ regions gives each region $m/R$, so a rare region's tail is protected only if $m/R \gtrsim 1/p_{\min}^{\text{region}}$.)*
5. **Recombination benefit** (Muller's ratchet): a single clonal lineage accumulates irreversible loss at the drift rate; the expected tail coverage of a pupil drawn from $K_T$ teachers with pairwise retention-correlation $\rho$ interpolates between the single-teacher value ($\rho = 1$) and the union of $K_T$ independent lineages ($\rho = 0$). With the construction of §2.7.1 (each teacher retains a fraction $q$ of the $T$ tail items, at exact pairwise correlation $\rho$) the interpolation is **closed-form**: the expected number of tail items retained by at least one teacher is
$$U(K_T,\rho,q) = T\left[\rho q + (1-\rho)\left(1-(1-q)^{K_T}\right)\right].$$
Check the limits: $K_T=1 \Rightarrow Tq$ (single teacher, independent of $\rho$); $\rho=1 \Rightarrow Tq$ (identical teachers, union = one); $\rho=0 \Rightarrow T[1-(1-q)^{K_T}]$ (independent teachers, maximal union). The pupil's realised tail coverage tracks $U$ *up to* the drift-survival threshold of prediction 4 — a tail item present in the mixture only survives the pupil's resampling of size $n$ if its mixture mass clears $\sim 1/n$, which ties E4 back to E2/E3. (Numerically verified against the construction to three decimals; it is a `test_scientific_validation.py` assertion.)
### 2.5 Experiments E1E6
Each experiment is one config file (§4), one runner invocation, one results artifact, and one figure script. Each states its prediction and its **falsifier** — the outcome that would refute the corresponding claim in the perspective paper.
**E1 — Reproduce collapse (null model).** $m=0$, single teacher, no selection. *Expect:* $H_t = H_0(1-1/n)^t$; support shrinks to 1; forward KL diverges; tail mass → 0, tail first. *Validates against:* predictions 12. *Falsifier of the harness (not the theory):* if drift does not reproduce the analytic decay, the simulator is wrong — fix before proceeding.
**E2 — Grounding phase boundary (headline).** Sweep $g = m/(n+m)$ from 0 to a high value; single teacher; uniform grounding; no selection. *Expect:* a critical $g^\*$ below which tail mass and $H$ decay to (near) zero and above which they stabilise at a positive stationary value tracking $H_{\text{eq}} = \theta/(1+\theta)$. *This is the paper's central quantitative prediction: the ratio of inherited-to-grounded information has a threshold.* *Falsifier:* if the stationary tail mass is flat in $g$, or if it only stabilises as $g \to 1$ (you always need essentially all-real data), then inheritance buys nothing and the multigenerational thesis is refuted. Report $g^\*$ with a CI.
**E3 — Region-matched grounding.** Fixed total $m$; compare `uniform` vs. `matched` allocation across $R$ regions, with one designated "inherited-but-not-freshly-grounded-under-uniform" region carrying a rare tail. *Expect:* under `uniform`, the target region's tail collapses even though global grounding is nonzero; under `matched`, it persists. *Validates:* prediction 4 and the "grounding must overlap the content it protects" claim. *Falsifier:* if uniform grounding protects the region as well as matched, the region-matching requirement is unnecessary and that paragraph of the paper should be cut.
**E4 — Multi-teacher decorrelation.** Teachers built by the §2.7.1 construction; sweep number of teachers $K_T \in \{1,2,3,5\}$ and retention-correlation $\rho \in [0,1]$ (at fixed marginal retention $q$); matched total data budget (so more teachers ≠ more data — each contributes $n/K_T$). Report **two** coverages: the construction-level union $U(K_T,\rho,q)$ (must match the §2.4-5 closed form exactly) and the post-distillation *surviving* coverage after the pupil's resampling. *Expect:* both rise with $K_T$ and with $(1-\rho)$; collapse suppression $\propto$ decorrelation; at $\rho=1$, multiple teachers give no benefit over one; and the gap between union and surviving coverage shrinks as grounding $g$ rises (recombination supplies the tail, grounding holds it). *Validates:* prediction 5 (now closed-form). *Falsifier:* if $K_T$ decorrelated teachers give no surviving-coverage benefit over one at matched budget, the recombination claim dies and single-teacher distillation is fine.
**E5 — Quality-diversity vs. greedy.** Same starting diversity; `greedy` vs. `qd` selection; sweep novelty exponent $\alpha$. *Expect:* `greedy` → fixation ($H \to 0$); `qd` holds $H$ at a positive plateau and re-introduces lost tail items. *Validates:* the two-level anti-convergence argument (§7 of paper). *Falsifier:* if `qd` does not maintain higher stationary $H$ than `greedy`, quality-diversity is not doing the work the paper assigns it.
**E6 — Re-minting gate (irreversibility).** Run a lineage to a chosen diversity level, then re-mint (freeze current $p_t$ as new reference, discard original $p^\*$); compare re-minting at high $H$ vs. low $H$. *Expect:* re-mint while collapsed → KL to *original* truth is locked high forever (tails unrecoverable); gated re-mint at high $H$ → no lock-in. *Validates:* §11's "re-minting is irreversible; gate it on diversity." *Falsifier:* if a collapsed lineage recovers its original-truth tails after re-minting, the irreversibility warning is overstated.
### 2.6 Layer-1 falsifiers, collected
The model is built to be *able to kill the thesis*. If E2 shows no threshold, or E4 shows no decorrelation benefit, or E6 shows no lock-in, the corresponding claims are refuted and the paper must say so. A blueprint that cannot fail is not a test. The single load-bearing positive result is E2's phase boundary at $g^\* \ll 1$: it says a little grounding protects a lot of inheritance, which is the whole economic and conceptual bet of the architecture.
### 2.7 Implementation spec for Layer 1
**Language / libraries.** Python ≥ 3.11; NumPy, SciPy (stats), pandas (results), matplotlib (figures). No GPU. No other heavy deps. Everything seedable from a single integer.
**Core module interfaces** (the coding agent should implement to these signatures; names are normative so downstream scripts are stable):
```python
# knowledge/truth.py
def make_true_distribution(K: int, R: int, tail: str, tail_frac: float,
zipf_s: float, seed: int) -> TrueDist:
"""Return p_star (length K), region assignment (length K), and the tail mask."""
# knowledge/teachers.py
def make_retention_matrix(T: int, K_T: int, rho: float, q: float, rng) -> np.ndarray:
"""Return an (K_T, T) binary retention matrix R with exact marginal retention
E[R]=q and exact pairwise column-correlation rho, via the shared-switch
construction of §2.7.1. rho=0 -> independent tails; rho=1 -> identical."""
def make_correlated_teachers(p_star, tail_mask, K_T: int, rho: float, q: float,
region_assignment=None, region_specialisation=False,
tail_floor: float = 1e-9, seed: int = 0) -> list[np.ndarray]:
"""Build K_T teacher distributions from a retention matrix (§2.7.1): every teacher
keeps all head items; teacher k keeps tail item j at ~p_star mass iff R[k,j]=1,
else at tail_floor; renormalise. If region_specialisation, force R[k,j]=1 for tail
items in teacher k's home region and apply the rho construction only off-home.
The exact-construction path is preferred for E4; a drift-based path (running
independent grounded lineages) is provided as a realism cross-check only."""
# knowledge/step.py
def generation_step(teachers, p_star_eff, cfg, rng) -> np.ndarray: ...
def apply_selection(p, p_star, mode: str, alpha: float) -> np.ndarray: ...
def structured_multinomial(m_vector, p_star, regions, policy: str, rng) -> np.ndarray: ...
# knowledge/lineage.py
def run_lineage(cfg, seed) -> pd.DataFrame:
"""Run T generations for one seed; return a tidy frame with one row per
(generation) and columns for every metric in §2.3 (global and per-region)."""
# knowledge/metrics.py
def forward_kl(p_star, p, eps): ...
def heterozygosity(p): ...
def tail_mass(p, tail_mask): ...
def support_size(p, eps): ...
# knowledge/experiment.py
def run_experiment(cfg) -> pd.DataFrame:
"""Sweep the declared parameter grid x n_replicates seeds; return long-form
results with confidence intervals; write parquet + the exact resolved config."""
```
**Config schema** (one YAML per experiment; all parameters explicit, no magic numbers in code). Illustrative default:
```yaml
experiment: E2_grounding_phase_boundary
seed: 20260704
n_replicates: 100
generations: 300
truth:
K: 1000 # number of knowledge items
R: 10 # regions
tail: zipf # {zipf, twocomponent}
zipf_s: 1.1
tail_frac: 0.5 # fraction of items designated 'tail'
tail_threshold: 1.0e-3
dynamics:
n: 200 # distillation sample size (drift strength)
teachers:
K_T: 1
rho: 0.0
grounding:
sweep: {param: g, values: [0.0, 0.01, 0.02, 0.05, 0.1, 0.2, 0.4]}
policy: uniform # {uniform, matched}
selection:
mode: none # {none, greedy, qd}
novelty_alpha: 0.0
remint:
enabled: false
period: null
H_gate: null
metrics:
kl_floor: 1.0e-9
output:
dir: results/fig2_grounding_sweep/
save_per_generation: true
```
**Determinism.** One master seed → per-replicate seeds via a documented `np.random.SeedSequence` spawn. No global RNG state; pass `rng` explicitly everywhere. Results are a pure function of the resolved config. The resolved config (after sweep expansion) is written next to the results artifact.
**Outputs.** Each experiment writes: `results.parquet` (long form), `resolved_config.yaml`, and a `manifest.json` recording library versions, seed, git commit, and a content hash of the results. Figures are generated by a separate `figures/plot_EX.py` reading only `results.parquet`, so any figure is regenerable from committed data without rerunning the simulation.
#### 2.7.1 The correlated-teacher construction (E4's one non-obvious piece)
E4 needs teacher sets whose pairwise **retention-correlation** $\rho$ is a clean, swept knob. Tuning the drift parameters until an emergent $\rho$ appears is the wrong design: that $\rho$ would be a tangled function of $n$, $m$, tail size and generation count, un-sweepable and confounded with the very drift whose effect E4 is trying to hold fixed. So we **construct the retention structure directly**, with $\rho$ an independent control. For an experiment whose entire purpose is to isolate the effect of decorrelation, $\rho$ must be manipulated, not inferred.
**The mechanism — shared-switch exchangeable Bernoulli.** Let the tail have $T$ items. Each teacher $k$ retains a fraction $q$ of them; we want the retention indicators to have marginal $q$ and *exact* pairwise correlation $\rho$ across teachers. For each tail item $j$:
1. draw a **shared switch** $z_j \sim \mathrm{Bernoulli}(\rho)$ (one draw, common to all teachers for item $j$);
2. draw a **shared retention** $s_j \sim \mathrm{Bernoulli}(q)$ (common to all teachers);
3. draw **independent retentions** $u^{(k)}_j \sim \mathrm{Bernoulli}(q)$, one per teacher;
4. set $r^{(k)}_j = s_j$ if $z_j = 1$, else $u^{(k)}_j$.
That is the whole construction. It yields a retention matrix $R \in \{0,1\}^{K_T \times T}$.
**Why it is exact.** Marginally $\mathbb{E}[r^{(k)}_j] = \rho q + (1-\rho)q = q$ regardless of $\rho$. For any two teachers, conditioning on the switch gives $\mathbb{E}[r^{(k)}_j r^{(k')}_j] = \rho\,\mathbb{E}[s_j^2] + (1-\rho)q^2 = \rho q + (1-\rho)q^2$ (using $s_j^2 = s_j$), so $\mathrm{Cov} = \rho q(1-q)$ and, since $\mathrm{Var} = q(1-q)$, the pairwise correlation is exactly $\rho$. The set is *exchangeable* — every teacher pair has the same $\rho$ — which is what makes $\rho$ a single scalar knob. (Verified numerically: marginal and pairwise correlation hit target across a $(\rho, q)$ grid, and the union closed form below matches to three decimals.)
**From retention to teacher distributions.** Given $R$, build teacher $k$'s distribution $p^{(k)}$: assign every **head** item its $p^\*$ mass (all teachers keep the common core); assign **tail** item $j$ its $p^\*_j$ mass if $r^{(k)}_j = 1$ and a floor $\varepsilon_{\text{tail}}$ otherwise; renormalise. (Renormalising lets the mass of dropped tails flow to what the teacher kept — the realistic signature of a partially-collapsed model concentrating on its survivors. Reallocating instead to the head is a documented config switch.)
**Region specialisation (structured decorrelation, optional).** With `region_specialisation=True`, give each teacher a **home region** and force $r^{(k)}_j = 1$ for every tail item $j$ in teacher $k$'s home region (each teacher fully retains the tails of the region it ground against reality), applying the $\rho$ construction only to off-home tail items. This models "each teacher ground a different region," ties E4 to E3's region-matched grounding, and is the discrete image of the perspective paper's "distil from teachers who each earned a different region."
**The analytic target (E4's exact check).** The probability that tail item $j$ is retained by at least one of $K_T$ teachers is $\rho q + (1-\rho)\big(1-(1-q)^{K_T}\big)$, so the expected **union tail-coverage** is
$$U(K_T,\rho,q) = T\left[\rho q + (1-\rho)\left(1-(1-q)^{K_T}\right)\right],$$
the closed form of §2.4-5. `make_retention_matrix` must reproduce it (and the target $\rho$, $q$) within Monte-Carlo tolerance in `test_scientific_validation.py`. The pupil's *realised* coverage after distillation equals $U$ only for tail items whose mixture mass clears the drift-survival threshold $\sim 1/n$ (§2.4-4) — so E4 should report both the union coverage (construction-level) and the post-distillation surviving coverage (dynamics-level), and their gap is itself informative: it is exactly the tail that recombination *supplied* but drift *re-erased* because grounding was too thin to hold it.
**Continuous-mass extension (optional realism, not a headline).** Binary retention is the default because it matches prediction 5 exactly. For a realism cross-check, replace binary retention with correlated *masses*: draw per-teacher log-masses on tail items from a multivariate normal with equicorrelation $\rho$ (a Gaussian copula), exponentiate, and normalise. This generalises to an arbitrary correlation *matrix* between teachers (unequal pairwise $\rho$), which the shared-switch construction — being exchangeable — cannot express; use it only if the paper later needs non-exchangeable teacher sets.
**Optional abstract treatment of the horizontal claim (§2.7-H, low priority).** If time permits, model a domain as a modular graph with tunable modularity $Q$; define the optimal number of concurrent specialists as the point where marginal coverage gain from an added specialist falls below a cost; show optimal specialist count rises with $Q$. This is the *horizontal* prediction in abstract form and does not require the neural layer. Flag clearly as exploratory; it is not a headline and can be dropped without weakening the paper.
---
## 3. Layer 2 — the neural existence proof
### 3.1 Purpose and the single objection it answers
Layer 1 assumes the tail-deletion operator (drift). A reviewer will say: *you built your conclusion into the operator.* Layer 2 answers exactly that objection and no more. It does not need to be a society. It needs to show, in **real LoRA-adapted weights**, that the *sign* of three effects is as Layer 1 predicts: dry inheritance degrades, grounded inheritance holds; multiple decorrelated teachers preserve capability that one teacher sheds; and across generations general capability holds or climbs while each specialty is re-earned and exceeded. If those three signs appear at 1B scale on one GPU, the abstraction in Layer 1 is grounded in mechanism.
Minimality is a virtue here, not a compromise. Every additional degree of freedom (bigger model, more generations, more domains) multiplies cost and reviewer surface without strengthening the core claim. Build the smallest thing that can show the signs.
### 3.2 Design choices (with open-science defaults)
**Base model.** Default to a *fully open* small model to honour reproducibility end-to-end: **OLMo-2-1B** or **SmolLM2-1.7B** (open weights, open or well-documented data, permissive licence). Capable fallback if the open models are too weak on the task: **Qwen2.5-1.5B-Instruct**. Pin the exact Hugging Face revision hash in config; never track `main`. The pipeline must be model-agnostic behind a thin adapter so swapping bases is a config change.
**Specialisation.** LoRA (via PEFT), small rank (e.g. 816), on a single task family = one "region." Hours on one consumer GPU, consistent with the perspective paper's cost claim. Each teacher is one base + one LoRA adapter.
**The verifier = "reality's no."** The domain must have a cheap, deterministic, uncontrollable oracle. **Program synthesis with unit tests** is ideal: a generated solution either passes its tests or it does not, and neither the model nor the experimenter controls the verdict. This is the minimal honest instance of the perspective paper's "predictive success under intervention." Execution happens in a sandbox (subprocess with a hard timeout and no network, run inside the container of §4; document the sandbox precisely).
**Task domain — synthetic-primary, benchmark-secondary.** To keep the Layer-1 abstractions (region, rarity, tail) exact, the *primary* domain is a **synthetic program-synthesis generator**: a family of small, verifiable tasks parameterised by *operation type* (the region: e.g. string ops, list ops, arithmetic, dict manipulation, recursion) and *difficulty*, with I/O unit tests generated automatically. Task-type frequency is tunable, so "rare tail task types" is a dial, exactly matching Layer 1's rarity. For *external validity*, add a secondary evaluation on a held-out slice of a public benchmark (e.g. MBPP-sanitised / HumanEval) — pinned by version — to show the effect is not an artefact of the synthetic generator. The synthetic generator's spec (grammar, per-region templates, test-generation rule, rarity distribution) is itself a committed, seeded artifact.
**Grounding, concretely.** A generational passage produces pupil training data as a mixture of (i) *inherited* teacher-generated solutions and (ii) *grounded* solutions that have been **filtered to pass the verifier** and/or freshly drawn from verified references in the target region. The **grounding fraction $g$** is the proportion of verifier-passed/real items in the pupil's training mixture — the *same knob* as Layer 1's $g$. "Dry" = $g$ low / unfiltered teacher output; "grounded" = $g$ raised with region-matched verified data.
**Recombination, concretely.** Multiple teachers = multiple LoRA specialists on *disjoint* task families (decorrelated by construction). The pupil is trained on the pooled outputs of all teachers (distillation) or, as a cheaper alternative, the specialists are **merged** (M2N2 / model-merge style, citing Sakana's demonstration that this search runs with no retraining) and the merged model is the pupil. Distillation and merging are two config-selectable recombination operators; report at least distillation, and merging if compute allows.
### 3.3 Contrasts C1C4 (the whole experimental content)
**C1 — Dry vs. grounded, single teacher, across generations.** One teacher, one region. Arm A: pupil trained on *unfiltered* teacher outputs (dry, $g\approx0$). Arm B: pupil trained on *verifier-passed* teacher outputs plus fresh verified region data ($g>0$). Run 23 generations. *Expect:* Arm A pass@k degrades generation over generation and its solution diversity narrows; Arm B holds. *Maps to:* Layer-1 E2. *Falsifier:* if dry inheritance does not degrade at this scale, collapse is not reachable here and the neural claim is unsupported (report honestly; possibly scale down grounding or up generations).
**C2 — One teacher vs. N complementary teachers, matched data budget.** Fix the pupil's total training-token budget. Arm A: all budget from one teacher (one region). Arm B: same budget split across $N$ teachers on disjoint regions. Evaluate the pupil on *all* regions, including rare-type tasks. *Expect:* Arm B retains capability across the union of regions (the "tail" of rare task types survives); Arm A loses regions it did not inherit. *Maps to:* Layer-1 E4. *Falsifier:* no union benefit at matched budget → recombination claim unsupported neurally.
**C3 — The vertical claim (load-bearing).** Over 23 generations, track (a) a *general* held-out benchmark spanning all regions and (b) *per-specialty* performance, where each generation re-specialises (re-earns) its region against the verifier. *Expect:* general benchmark is **monotonically non-decreasing** across generations, while each specialty is re-earned and its peak **exceeds the parent's** peak in that specialty. *This is the test of the actual thesis, not the borrowed Shumailov scaffolding.* *Maps to:* the paper's vertical prediction. *Falsifier:* if general capability falls across generations, or specialties are not re-earned-and-exceeded, the multigenerational ratchet is not demonstrated.
**C4 — Recombination operator: distillation vs. merging (optional).** If compute allows, show C2's benefit is robust to whether recombination is done by distillation or by weight-merging. Strengthens generality and directly connects to the M2N2 citation. Drop first if time-constrained.
### 3.4 Metrics (Layer 2)
- **pass@1 and pass@k** on held-out tasks, per region and overall (the neural analogue of KL-to-truth / capability).
- **Region coverage** = fraction of regions (incl. rare task types) with pass@1 above a floor — the neural analogue of *support size* / tail survival.
- **Solution diversity** = distinct-$n$ or behavioural diversity of generated solutions per task (the neural analogue of heterozygosity $H$; narrowing diversity is the neural signature of collapse).
- **Generational degradation curve** = each metric as a function of generation index, per arm.
All with multiple seeds (fewer than Layer 1 — GPU cost — but at least 3; report per-seed points, not just means, given small $n$). Pin decoding parameters (temperature, top-p, max tokens, sampling seed) in config; they materially affect pass@k and must not float.
### 3.5 Layer-2 honesty riders
GPU non-determinism means Layer 2 is *statistically* reproducible (same distribution of outcomes under re-run), not bitwise reproducible. Document this explicitly; pin everything pinnable (model revision, dataset version, decoding params, library versions, seeds); report seeds individually. The claim Layer 2 supports is directional ("the sign of the effect is as predicted"), and the writing must not overclaim precision the setup cannot deliver.
### 3.6 Implementation spec for Layer 2
**Libraries.** PyTorch; Hugging Face `transformers` + `peft` (LoRA); `datasets`; optional `vllm` for fast generation; the synthetic task generator (project-local). Execution sandbox: `subprocess` with `resource` limits and timeout, inside the container.
**Module interfaces (normative names):**
```python
# neural/tasks.py
def generate_task_bank(regions, rarity, n_tasks, seed) -> TaskBank:
"""Synthetic verifiable tasks; each task carries prompt, region, difficulty,
and an executable unit-test suite."""
def verify(solution_code: str, task) -> VerifyResult:
"""Run tests in a sandbox; return pass/fail + diagnostics. Deterministic."""
# neural/specialise.py
def train_lora_specialist(base_id, revision, region, task_bank, lora_cfg, seed) -> AdapterPath: ...
# neural/distill.py
def generate_teacher_data(teachers, task_bank, grounding_fraction, policy, decode_cfg, seed) -> Corpus:
"""Produce the pupil's training corpus: mixture of inherited (teacher) and
grounded (verifier-passed / fresh-verified) items, region-matched."""
def train_pupil(base_id, revision, corpus, lora_cfg, seed) -> AdapterPath: ...
# neural/merge.py
def merge_specialists(base_id, adapters, method, seed) -> ModelPath: # optional (C4)
# neural/evaluate.py
def evaluate(model, task_bank_heldout, decode_cfg, seed) -> pd.DataFrame:
"""pass@k, per-region coverage, solution diversity; tidy per-(region) frame."""
# neural/generation_loop.py
def run_generations(cfg, seed) -> pd.DataFrame:
"""Orchestrate T generations for one arm; log every metric per generation."""
```
**Config** mirrors Layer 1's structure (one YAML per contrast; pinned `base_id` + `revision`; explicit `grounding_fraction`, `n_teachers`, `regions`, `generations`, `decode_cfg`, `lora_cfg`, `seed`, `n_seeds`). Same output contract: `results.parquet` + `resolved_config.yaml` + `manifest.json` (with model revision hashes and dataset versions). Figures regenerable from `results.parquet` alone.
---
## 4. Reproducibility and engineering standard (both layers)
Open science is a hard requirement of this project, not a preference. The standard below is normative.
**Environment.** Pin everything. Provide (a) a `pyproject.toml` + lockfile via **uv** (fast, reproducible resolver) and (b) an **Apptainer/Singularity** definition file (HPC-friendly, rootless, open) that builds the exact environment; optionally a Dockerfile. The container is the source of truth for "it runs." Record Python, CUDA, and key library versions in every run's `manifest.json`.
**Seeding.** One master seed per experiment in config; derive all sub-seeds via `SeedSequence.spawn`; never touch global RNG state. Layer 1 is bitwise-reproducible from seed. Layer 2 is statistically reproducible; document the residual GPU non-determinism and set the available determinism flags (`torch.use_deterministic_algorithms(True)` where feasible, cudnn deterministic, documented exceptions).
**Configuration.** No magic numbers in code — every parameter lives in a YAML resolved at run time; the *resolved* config is written beside results. Use a single config system (Hydra or a thin equivalent). Sweeps are declared in config, not hard-coded in scripts.
**Data & model provenance.** Layer 1 data is synthetic-from-seed (fully reproducible; no external data). Layer 2 pins model revision hashes and dataset versions; the synthetic task generator is committed and seeded. Cache external downloads with recorded hashes.
**Experiment tracking.** Prefer open tooling: **MLflow** (open source) or plain versioned CSV/Parquet + committed configs; avoid closed SaaS trackers to keep the pipeline fully open. Whatever is chosen, the invariant is: every figure is a pure function of a committed results artifact.
**Testing.** `pytest`. Two kinds of test, and both are required:
- *Correctness tests*: the module does what it says (shapes, normalisation, sandbox isolation).
- *Scientific validation tests*: the simulator reproduces the §2.4 analytic results within tolerance (heterozygosity decay, fixation probability, mutationdrift equilibrium). These tests failing means the science is wrong, not just the code — they are the spine of trust in Layer 1.
**Automation.** A `Makefile` (or `justfile`) with targets: `env`, `test`, `layer1` (runs E1E6), `layer2` (runs C1C3, C4 optional), `figures`, `paper` (assembles the figure manifest), `all`, `clean`. One command reproduces the study from a clean checkout inside the container.
**Repro entry point.** A top-level `reproduce.sh` that: builds/enters the container, runs tests, runs all experiments at the committed seeds, regenerates all figures, and writes a `REPRODUCED.md` diff against committed result hashes. If hashes match (Layer 1) / distributions match within CI (Layer 2), the run is verified.
**Licensing & citation.** Author to choose; suggested: code under a permissive OSI licence (MIT/Apache-2.0) or copyleft (GPL-3.0) per the author's open-source preference; text/figures under CC-BY. Include `LICENSE`, `CITATION.cff`, and a `DATA_AND_MODELS.md` recording every external artifact and its pinned version/hash.
---
## 5. Repository layout
```
lamarckian-society/
├── README.md # what this is, how to reproduce
├── reproduce.sh # one-command full reproduction (in-container)
├── Makefile # env, test, layer1, layer2, figures, all
├── pyproject.toml # deps
├── uv.lock # pinned resolution
├── apptainer.def # container definition (source of truth)
├── Dockerfile # optional
├── LICENSE CITATION.cff DATA_AND_MODELS.md
├── configs/
│ ├── layer1/E1..E6.yaml
│ └── layer2/C1..C4.yaml
├── src/
│ ├── knowledge/ # Layer 1: truth, teachers, step, lineage, metrics, experiment
│ └── neural/ # Layer 2: tasks, specialise, distill, merge, evaluate, generation_loop
├── figures/ # plot_EX.py / plot_CX.py — read results.parquet only
├── results/ # written artifacts (gitignored, hashes tracked)
├── tests/
│ ├── test_correctness.py
│ └── test_scientific_validation.py # §2.4 analytic checks
└── paper/
├── blueprint.md # this document
└── figure_manifest.md # claim -> experiment -> figure
```
---
## 6. Traceability: claim → experiment → figure
Every claim the technical paper will make is bound to exactly one experiment and one figure. This table is the paper's spine and the coding agent's definition of done.
| Perspective-paper claim | Layer | Experiment | Primary figure | Analytic check | Falsifier |
|---|---|---|---|---|---|
| Distillation without grounding collapses, tail first, at rate set by $n$ | 1 | E1 | $H_t$ decay + tail mass vs. gen | Pred. 12 | harness invalid if decay ≠ analytic |
| A critical grounding ratio $g^\*$ separates ratchet from collapse | 1 | E2 | tail mass / $H$ stationary vs. $g$ | Pred. 3 | flat in $g$, or only stable at $g\to1$ |
| Grounding must be region-matched, not merely present | 1 | E3 | target-region tail: uniform vs. matched | Pred. 4 | uniform protects as well as matched |
| Multi-teacher distillation suppresses collapse ∝ decorrelation | 1 | E4 | tail coverage surface over $(K_T,\rho)$ | Pred. 5 (closed form $U(K_T,\rho,q)$) | no benefit at matched budget |
| QD selection maintains diversity where greedy fixes it | 1 | E5 | $H$ plateau: qd vs. greedy | (mutation-selection-drift) | qd ≤ greedy in stationary $H$ |
| Re-minting is irreversible; gate on diversity | 1 | E6 | KL-to-original vs. $H$-at-remint | — | collapsed lineage recovers after remint |
| Dry inheritance degrades in real weights; grounded holds | 2 | C1 | pass@k vs. gen, dry vs. grounded | — | dry does not degrade |
| Complementary teachers preserve capability one teacher sheds | 2 | C2 | region coverage, 1 vs. N, matched budget | — | no union benefit |
| **General knowledge climbs while each specialty is re-earned and exceeded** | 2 | C3 | general benchmark ↑ + per-specialty re-earn | — | general falls / no re-earn |
| Recombination benefit robust to distillation vs. merging | 2 | C4 (opt) | C2 metric under both operators | — | benefit only under one operator |
---
## 7. Suggested build order for the coding agent
Staged so that each step is independently testable and the cheapest, highest-value results land first. Do not start Layer 2 until Layer 1's scientific-validation tests pass.
1. **Scaffold + environment.** Repo layout (§5), container (§4), `pytest` skeleton, config system, seeding utilities. Target `make test` green on trivial tests.
2. **Layer 1 core + validation.** `knowledge/` modules to the §2.7 interfaces. Implement the null model first. Write and pass `test_scientific_validation.py` against §2.4 predictions 12. **Gate: do not proceed until drift matches analytic decay.**
3. **Layer 1 mechanisms + E1E2.** Add grounding; validate mutationdrift equilibrium (pred. 3); run E1 and the E2 phase-boundary sweep; produce the headline figure. This is the paper's core result and it should exist before anything neural.
4. **Layer 1 E3E6.** Region-matching, multi-teacher/decorrelation, QD-vs-greedy, re-minting gate, with figures. Layer 1 is now a complete, laptop-reproducible paper on its own.
5. **Layer 2 scaffold + verifier.** Synthetic task generator, sandboxed `verify`, evaluation harness. Test the verifier's determinism and isolation before any training.
6. **Layer 2 C1 + C3.** Single-teacher dry-vs-grounded (C1) and the vertical claim (C3) — the two that most directly test the thesis. C3 is load-bearing; prioritise it.
7. **Layer 2 C2 (+ C4 if compute allows).** Multi-teacher recombination; optional merging operator.
8. **Reproduction pass.** `reproduce.sh` end-to-end; commit result hashes; write `REPRODUCED.md`; assemble the figure manifest.
**Definition of done:** every row of §6 has a committed figure produced by `make figures` from committed results, every §2.4 analytic check passes in CI, and `reproduce.sh` verifies from a clean checkout inside the container.
---
*End of blueprint v1. The perspective paper states the idea; this document states the test. If Layer 1's E2 finds no threshold, or E4 no decorrelation benefit, or Layer 2's C3 no vertical climb, the thesis is wrong in exactly the places these experiments probe — which is the point of writing them down this precisely.*

View file

@ -1,291 +0,0 @@
"""Build a Zotero-importable library from the manuscript's reference list.
For each of the numbered references in paper/manuscript/main.md: take the DOI printed in the entry when
there is one, otherwise ask Crossref for it by title (accepting only a high-scoring match whose title
really is the same, checked by normalised comparison). Then fetch authoritative metadata for every
resolved DOI by content negotiation against doi.org, which serves Crossref and DataCite alike, and
write the result as CSL-JSON plus RIS.
Entries whose DOI cannot be resolved (pre-DOI literature, books, chapters) are reported and written
from the manuscript's own metadata so nothing is silently dropped.
Usage: python paper/manuscript/build_zotero_library.py
"""
from __future__ import annotations
import json
import re
import sys
import time
import urllib.parse
import urllib.request
from pathlib import Path
MAIN = Path(__file__).resolve().parent / "main.md"
OUT = Path(__file__).resolve().parent / "refs"
MAILTO = "g.gilestro@imperial.ac.uk" # Crossref polite pool
UA = f"LamarckianAI-refs/1.0 (mailto:{MAILTO})"
# Reference numbers whose sources predate DOIs or are books/chapters: never send these to Crossref
# title search, because it returns confident nonsense for them.
NO_DOI_EXPECTED = {33, 35, 39} # Jenkin 1867; Fisher 1930 (book); Templeton 1986 (chapter)
# DOIs the title search could not find and that were verified by hand against the publisher record.
DOI_OVERRIDE = {
17: "10.1038/s41562-023-01742-2", # Brinkmann et al., Machine culture (Nat. Hum. Behav.)
64: "10.48550/arXiv.1805.06370", # Schwarz et al., Progress & Compress (no Crossref DOI)
}
# The three genuinely pre-DOI sources, written out rather than parsed, so the Zotero records are
# complete instead of merely non-empty.
HAND_WRITTEN = {
33: {"type": "article-journal", "title": "[Review of] The Origin of Species",
"author": [{"given": "Fleeming", "family": "Jenkin"}],
"container-title": "The North British Review", "volume": "46", "page": "277-318",
"issued": {"date-parts": [[1867]]}},
35: {"type": "book", "title": "The Genetical Theory of Natural Selection",
"author": [{"given": "Ronald A.", "family": "Fisher"}],
"publisher": "Clarendon Press", "publisher-place": "Oxford",
"issued": {"date-parts": [[1930]]}},
39: {"type": "chapter", "title": "Coadaptation and outbreeding depression",
"author": [{"given": "Alan R.", "family": "Templeton"}],
"editor": [{"given": "Michael E.", "family": "Soulé"}],
"container-title": "Conservation Biology: The Science of Scarcity and Diversity",
"publisher": "Sinauer Associates", "publisher-place": "Sunderland, MA",
"page": "105-116", "issued": {"date-parts": [[1986]]}},
}
def get(url: str, accept: str | None = None, tries: int = 3) -> bytes | None:
req = urllib.request.Request(url, headers={"User-Agent": UA})
if accept:
req.add_header("Accept", accept)
for i in range(tries):
try:
with urllib.request.urlopen(req, timeout=30) as r:
return r.read()
except Exception as e: # noqa: BLE001
if i == tries - 1:
print(f" ! {type(e).__name__}: {str(e)[:80]}", file=sys.stderr)
time.sleep(1.5 * (i + 1))
return None
def parse_refs() -> list[tuple[int, str]]:
refs = MAIN.read_text().split("## References")[1]
out = []
for line in refs.splitlines():
if m := re.match(r"^(\d+)\. (.*)$", line):
out.append((int(m.group(1)), m.group(2).strip()))
return out
def strip_md(s: str) -> str:
return re.sub(r"[*_`]", "", s)
def guess_title(entry: str) -> str:
"""The title is the run of text between the author list and the italic venue or the year."""
t = strip_md(entry)
t = re.sub(r"\s*https?://\S+$", "", t).strip()
# drop the leading author list: everything up to the last ", " before the title is unreliable,
# so instead cut after the first ", " that follows an initial-style name block
m = re.match(r"^((?:[A-ZÀ-Þ]\.\s*)+[^,]+,\s*)+", t)
rest = t[m.end():] if m else t
rest = re.sub(r"^et al\.,\s*", "", rest)
# the title ends at the venue (". *Venue*") or at " arXiv [Preprint]" or " (Year)"
rest = re.split(r"\.\s+(?:arXiv \[Preprint\]|[A-Z][a-zA-Z.\s&]*\*|Proc\.|Int\.|Adv\.|Conf\.|Nat\.|Trans\.)", rest)[0]
rest = re.split(r"\s*\(\d{4}\)", rest)[0]
return rest.strip(" .,")
def norm(s: str) -> str:
return re.sub(r"[^a-z0-9]", "", s.lower())
def crossref_by_title(title: str, year: str | None) -> tuple[str | None, str]:
q = urllib.parse.urlencode({"query.bibliographic": title, "rows": 5, "mailto": MAILTO})
raw = get(f"https://api.crossref.org/works?{q}")
if not raw:
return None, "crossref unreachable"
items = json.loads(raw).get("message", {}).get("items", [])
tn = norm(title)
for it in items:
cand = (it.get("title") or [""])[0]
cn = norm(cand)
if not cn:
continue
# accept only a genuine title match, not merely a high Crossref score
if cn.startswith(tn[:60]) or tn.startswith(cn[:60]):
return it.get("DOI"), f"matched: {cand[:70]}"
return None, f"no title match (best: {(items[0].get('title') or [''])[0][:60] if items else '-'})"
def csl_from_doi(doi: str) -> dict | None:
raw = get(f"https://doi.org/{urllib.parse.quote(doi)}",
accept="application/vnd.citationstyles.csl+json")
if not raw:
return None
try:
return json.loads(raw)
except json.JSONDecodeError:
return None
# ---------------------------------------------------------------- fallback CSL from the manuscript
def manual_csl(num: int, entry: str) -> dict:
t = strip_md(entry)
year = (re.search(r"\((\d{4})\)", t) or re.search(r"(\d{4})", t))
authors = []
m = re.match(r"^((?:[A-ZÀ-Þ]\.(?:\s*[A-ZÀ-Þ]\.)*\s+[^,]+,\s*)+)", t)
if m:
for name in re.findall(r"([A-ZÀ-Þ]\.(?:\s*[A-ZÀ-Þ]\.)*)\s+([^,]+)", m.group(1)):
authors.append({"given": name[0].strip(), "family": name[1].strip()})
venue = re.search(r"\*([^*]+)\*", entry)
vol = re.search(r"\*\*(\d+)\*\*", entry)
pages = re.search(r"\*\*\d+\*\*,\s*([\d\-]+)", entry)
return {k: v for k, v in {
"id": f"ref{num}",
"type": "book" if "Press)" in t or "Sinauer" in t else "article-journal",
"title": guess_title(entry),
"author": authors or None,
"container-title": venue.group(1) if venue else None,
"volume": vol.group(1) if vol else None,
"page": pages.group(1).replace("", "-") if pages else None,
"issued": {"date-parts": [[int(year.group(1))]]} if year else None,
"note": f"manuscript reference {num}; no DOI",
}.items() if v is not None}
def clean_text(s: str) -> str:
"""Publisher abstracts arrive with JATS tags, HTML entities, and hard line breaks; RIS is a
line-oriented format, so every field has to end up as one clean line."""
import html
s = re.sub(r"<[^>]+>", " ", s) # JATS/HTML tags
s = html.unescape(s)
return re.sub(r"\s+", " ", s).strip()
def clean_csl(c: dict) -> dict:
for k, v in list(c.items()):
if isinstance(v, str):
c[k] = clean_text(v)
elif isinstance(v, list) and v and isinstance(v[0], str):
c[k] = [clean_text(x) for x in v]
doi = c.get("DOI", "")
if doi.lower().startswith("10.48550/arxiv."):
# DataCite returns these uppercased and with no venue; restore the canonical DOI casing and
# give Zotero something to show in the publication field instead of a blank.
arxiv_id = doi.split(".", 2)[-1]
c["DOI"] = f"10.48550/arXiv.{arxiv_id}"
c["container-title"] = "arXiv"
c["number"] = f"arXiv:{arxiv_id}"
c["genre"] = "preprint"
return c
# Crossref reports its own type vocabulary alongside real CSL types; map both.
CSL2RIS_EXTRA = {"journal-article": "JOUR", "book-chapter": "CHAP", "proceedings-article": "CPAPER",
"posted-content": "JOUR", "book-section": "CHAP", "monograph": "BOOK"}
CSL2RIS = {"article-journal": "JOUR", "paper-conference": "CPAPER", "chapter": "CHAP",
"book": "BOOK", "article": "JOUR", "posted-content": "JOUR", "report": "RPRT",
"dataset": "DATA", "thesis": "THES"}
def ris_type(c: dict) -> str:
t = c.get("type", "")
return CSL2RIS.get(t) or CSL2RIS_EXTRA.get(t) or "JOUR"
def to_ris(c: dict, num: int) -> str:
L = [f"TY - {ris_type(c)}"]
for a in c.get("author") or []:
fam, giv = a.get("family", ""), a.get("given", "")
L.append(f"AU - {fam}, {giv}".rstrip(", ") if fam else f"AU - {a.get('literal', '')}")
ttl = c.get("title")
if isinstance(ttl, list):
ttl = ttl[0]
if ttl:
L.append(f"TI - {ttl}")
ct = c.get("container-title")
if isinstance(ct, list):
ct = ct[0] if ct else None
if ct:
L.append(f"{'BT' if ris_type(c) == 'CHAP' else 'T2'} - {ct}")
for ed in c.get("editor") or []:
L.append(f"A2 - {ed.get('family', '')}, {ed.get('given', '')}".rstrip(", "))
if c.get("number"):
L.append(f"AN - {c['number']}")
if c.get("publisher-place"):
L.append(f"CY - {c['publisher-place']}")
parts = (c.get("issued") or {}).get("date-parts") or [[]]
if parts and parts[0]:
L.append(f"PY - {parts[0][0]}")
for key, tag in (("volume", "VL"), ("issue", "IS"), ("publisher", "PB"), ("DOI", "DO"),
("URL", "UR"), ("abstract", "AB")):
if c.get(key):
L.append(f"{tag} - {c[key]}")
if c.get("page"):
pg = str(c["page"]).replace("", "-").split("-")
L.append(f"SP - {pg[0]}")
if len(pg) > 1:
L.append(f"EP - {pg[-1]}")
L.append(f"N1 - {c.get('note') or f'Manuscript reference {num}'}")
L.append("ER - \n")
return "\n".join(L)
def main() -> int:
refs = parse_refs()
print(f"{len(refs)} references parsed\n")
csls, report = [], []
for num, entry in refs:
doi = None
if num in HAND_WRITTEN:
c = dict(HAND_WRITTEN[num], id=f"ref{num}", note=f"Manuscript reference {num}; predates DOIs")
csls.append(c)
report.append((num, "HAND (pre-DOI source)", c["title"][:64], "written by hand"))
print(f" {num:3d} {'HAND (pre-DOI source)':52s} {c['title'][:56]}")
continue
if num in DOI_OVERRIDE:
doi, src = DOI_OVERRIDE[num], "verified by hand"
elif m := re.search(r"doi\.org/(10\.\S+?)\.?$", entry):
doi = m.group(1)
src = "in manuscript"
elif num not in NO_DOI_EXPECTED:
title = guess_title(entry)
yr = re.search(r"\((\d{4})\)", entry)
doi, why = crossref_by_title(title, yr.group(1) if yr else None)
src = f"crossref ({why})"
time.sleep(0.3)
else:
src = "pre-DOI / book — not searched"
c = csl_from_doi(doi) if doi else None
if c:
c["id"] = f"ref{num}"
c["note"] = f"Manuscript reference {num}"
status = f"OK {doi}"
else:
c = manual_csl(num, entry)
status = f"MANUAL ({src})" if not doi else f"MANUAL (DOI {doi} would not resolve)"
csls.append(c)
report.append((num, status, (c.get('title') or '')[:64], src))
print(f" {num:3d} {status:52s} {(c.get('title') or '')[:56]}")
time.sleep(0.2)
csls = [clean_csl(c) for c in csls]
(OUT / "references.json").write_text(json.dumps(csls, indent=1, ensure_ascii=False))
(OUT / "references.ris").write_text("".join(to_ris(c, n) for (n, _), c in zip(refs, csls)))
ok = sum(1 for _, s, _, _ in report if s.startswith("OK"))
print(f"\nresolved from DOI: {ok}/{len(refs)} manual: {len(refs)-ok}")
(OUT / "report.txt").write_text("\n".join(f"{n}\t{s}\t{t}\t{src}" for n, s, t, src in report))
return 0
if __name__ == "__main__":
sys.exit(main())

View file

@ -1,39 +0,0 @@
Giorgio F. Gilestro
Department of Life Sciences, Imperial College London
giorgio@gilest.ro
[Date]
Dear Editor,
Please consider the enclosed manuscript, "The evolution of sex for artificial intelligence: a population-genetic framework for multigenerational model populations", for publication as an Article in *Nature Machine Intelligence*.
Machine learning has become a population process. Public repositories hold millions of models, most of them fine-tunes, distillations or weight merges of a few ancestors; models learn from the output of earlier models; and merging, now mainstream practice with standard tooling, is described in its own literature with the words crossover, mutation and mate choice. A population whose members inherit from one another, recombine and retransmit is an evolving population in the technical sense, and the branch of biology built for that situation is the population genetics of sexual reproduction. That training on model output is genetic drift, with model collapse as its signature, has been established several times over. This paper takes the next step and develops the mechanisms population genetics offers for sustaining a population against drift (immigration, recombination, selection, population structure), and the point where they fail (reproductive isolation), and tests each of them in a chain from closed forms to trained networks to language models.
Four measurements are new, and each was chosen because the existing experimental designs could not make it.
First, a six-generation population of language models in which three lineages each learn a new skill every generation and then choose whether, and with whom, to merge. Merging has been iterated before, in evolutionary pools of fixed parents and in continual streams folded into one model, but never while the lineages were also learning. The population shows that obligate merging collapses once partners hold conflicting conventions (accuracy 0.65 to 0.27), that a merge each lineage may decline, or a fixed early stop, avoids the collapse at no cost against never merging, and that merging with one's own ancestor is safer than merging with a contemporary in every seed. A second curriculum decoupling partner complementarity from generation shows that declines track generation, which corrects an interpretation the first curriculum invited.
Second, model speciation as a named and tested question. Using the permutation-and-rescaling alignment of Git Re-Basin and REPAIR, the merge barrier between networks is separated into the part alignment removes and the part it cannot. Conflicting label maps leave a residual alignment does not touch, while six times the base training on non-conflicting tasks produces no isolation at all and the strongest rescue-by-merging in the paper, against the expectation that specialisation by itself erodes mergeability.
Third, a pre-merge predictive test on 39 language-model parent pairs across three decorrelated axes (conflict, compatible overlap, duration). Functional disagreement between parents predicts merge damage out of sample where LoRA-weight cosine and distance do not, in agreement with recent correlational reports. The control that matters is new: on a grid that varies conflict and shared training data together, weight cosine is the best predictor (ρ = 0.60), and adding overlap without conflict collapses it to 0.03. Any weight-geometry predictor validated on such a grid is reading the shared data, which bears on the merge-prediction literature independently of the biology.
Fourth, a conservation law for blending inheritance. Refitting a child on the average of several parents' outputs carries a rare capability across a generation no better than inheriting from one parent, to first order, so the gain of having several parents is realised only by operators that keep each parent's strongest contribution. The law fixes the null against which every recombination operator is judged and predicted the headroom rule measured in language models at two scales: routing and offspring selection beat the weight average wherever that average falls short of attainable performance (hard tasks at 7B, every seed), and add nothing where it does not.
Around these sit results that place the framework in the existing literature: a closed-form grounding equilibrium and per-item floor that agree with the fresh-data stability theorems and with the finding that absolute real-sample counts matter more than proportions; the transfer of every drift sign to trained networks with a measured, architecture-specific estimator bias; and a four-arm ablation of a composed population. Two refinements the framework proposed were not supported, and the paper says so.
I am submitting to *Nature Machine Intelligence* because the readers who make the decisions this paper prices (how much verified data a synthetic pipeline needs, whether to merge or route, when to stop merging, how to detect an incompatible pair before paying for the merge) are this journal's readers, and because the journal has already published evolutionary model merging as a research direction (Akiba et al., 2025). The paper gives that direction its theory and its failure modes. What biology receives in return is a model system where every genotype, environment and mating decision is observable and manipulable, so the paper should also interest the evolutionary biologists among your readership.
All code, configurations, seeds, results artefacts and a one-command reproduction script will be deposited openly with an archived DOI on publication; every figure regenerates from committed artefacts without re-simulation. The manuscript is not under consideration elsewhere and has not been published in any form. [A preprint has been / will be posted to arXiv.] I am the sole author and declare no competing interests.
Suggested referees:
- [Name, affiliation, email] (model merging)
- [Name, affiliation, email] (model collapse / synthetic data theory)
- [Name, affiliation, email] (population genetics of recombination and speciation)
- [Name, affiliation, email] (continual learning)
Excluded referees: [none / names].
Yours sincerely,
Giorgio F. Gilestro

Binary file not shown.

Binary file not shown.

Binary file not shown.

Binary file not shown.

Binary file not shown.

Binary file not shown.

Binary file not shown.

Binary file not shown.

Binary file not shown.

View file

@ -1,80 +0,0 @@
1 OK 10.48550/arXiv.2508.06811 Anatomy of a Machine Learning Ecosystem: 2 Million Models on Hug in manuscript
2 OK 10.48550/arXiv.2405.18432 Unsupervised Model Tree Heritage Recovery in manuscript
3 OK 10.48550/arXiv.2402.00699 PeaTMOSS: A Dataset and Initial Analysis of Pre-Trained Models i in manuscript
4 OK 10.48550/arXiv.2306.01708 TIES-Merging: Resolving Interference When Merging Models in manuscript
5 OK 10.1038/s42256-024-00975-8 Evolutionary optimization of model merging recipes crossref (matched: Evolutionary optimization of model merging recipes)
6 OK 10.48550/arXiv.2403.13257 Arcee's MergeKit: A Toolkit for Merging Large Language Models in manuscript
7 OK 10.48550/arXiv.2408.07666 Model Merging in LLMs, MLLMs, and Beyond: Methods, Theories, App in manuscript
8 OK 10.48550/arXiv.2503.01155 Nature-Inspired Population-Based Evolution of Large Language Mod in manuscript
9 OK 10.48550/arXiv.2508.16204 Competition and Attraction Improve Model Fusion in manuscript
10 OK 10.48550/arXiv.2501.05707 Multiagent Finetuning: Self Improvement with Diverse Reasoning C in manuscript
11 OK 10.48550/arXiv.2406.11704 Nemotron-4 340B Technical Report in manuscript
12 OK 10.48550/arXiv.2412.08905 Phi-4 Technical Report in manuscript
13 OK 10.48550/arXiv.2212.10560 Self-Instruct: Aligning Language Models with Self-Generated Inst in manuscript
14 OK 10.48550/arXiv.2401.05749 A Shocking Amount of the Web is Machine Translated: Insights fro in manuscript
15 OK 10.48550/arXiv.2403.07183 Monitoring AI-Modified Content at Scale: A Case Study on the Imp in manuscript
16 OK 10.48550/arXiv.2211.04325 Will we run out of data? Limits of LLM scaling based on human-ge in manuscript
17 OK 10.1038/s41562-023-01742-2 Machine culture verified by hand
18 OK 10.48550/arXiv.2304.03442 Generative Agents: Interactive Simulacra of Human Behavior in manuscript
19 OK 10.48550/arXiv.2402.01680 Large Language Model based Multi-Agents: A Survey of Progress an in manuscript
20 OK 10.48550/arXiv.2509.10147 Virtual Agent Economies in manuscript
21 OK 10.1038/s41586-024-07566-y AI models collapse when trained on recursively generated data crossref (matched: AI models collapse when trained on recursively generated data)
22 OK 10.1371/journal.pcbi.1002510 Structural Drift: The Population Dynamics of Sequential Learning crossref (matched: Structural Drift: The Population Dynamics of Sequential Learning)
23 OK 10.48550/arXiv.2604.08554 Drift and selection in LLM text ecosystems in manuscript
24 OK 10.48550/arXiv.2509.20101 First-Extinction Law for Resampling Processes in manuscript
25 OK 10.48550/arXiv.2407.17493 Model Collapse in the Self-Consuming Chain of Diffusion Finetuni in manuscript
26 OK 10.1016/s0079-7421(08)60536-8 Catastrophic Interference in Connectionist Networks: The Sequent crossref (matched: Catastrophic Interference in Connectionist Networks: The Sequential Le)
27 OK 10.1016/s1364-6613(99)01294-2 Catastrophic forgetting in connectionist networks crossref (matched: Catastrophic forgetting in connectionist networks)
28 OK 10.1016/0027-5107(64)90047-8 The relation of recombination to mutational advance crossref (matched: The relation of recombination to mutational advance)
29 OK 10.48550/arXiv.2510.16657 Escaping Model Collapse via Synthetic Data Verification: Near-te in manuscript
30 OK 10.48550/arXiv.2404.01413 Is Model Collapse Inevitable? Breaking the Curse of Recursion by in manuscript
31 OK 10.1093/genetics/16.2.97 EVOLUTION IN MENDELIAN POPULATIONS crossref (matched: EVOLUTION IN MENDELIAN POPULATIONS)
32 OK 10.1046/j.1523-1739.1996.10061509.x The OneMigrantperGeneration Rule in Conservation and Manageme crossref (matched: The OneMigrantperGeneration Rule in Conservation and Management)
33 HAND (pre-DOI source) [Review of] The Origin of Species written by hand
34 OK 10.48550/arXiv.2411.02207 Collective Model Intelligence Requires Compatible Specialization in manuscript
35 HAND (pre-DOI source) The Genetical Theory of Natural Selection written by hand
36 OK 10.1086/280418 Some Genetic Aspects of Sex crossref (matched: Some Genetic Aspects of Sex)
37 OK 10.48550/arXiv.2106.09685 LoRA: Low-Rank Adaptation of Large Language Models in manuscript
38 OK 10.1016/s0022-5193(87)80029-2 Towards a general theory of adaptive walks on rugged landscapes crossref (matched: Towards a general theory of adaptive walks on rugged landscapes)
39 HAND (pre-DOI source) Coadaptation and outbreeding depression written by hand
40 OK 10.1162/evco_a_00025 Abandoning Objectives: Evolution Through the Search for Novelty crossref (matched: Abandoning Objectives: Evolution Through the Search for Novelty Alone)
41 OK 10.48550/arXiv.2503.05683 WikiBigEdit: Understanding the Limits of Lifelong Knowledge Edit in manuscript
42 OK 10.48550/arXiv.2502.04390 In Praise of Stubbornness: An Empirical Case for Cognitive-Disso in manuscript
43 OK 10.48550/arXiv.2607.09202 Interference and Retention in Continual Learning in manuscript
44 OK 10.1017/s0016672300033140 A general model for the evolution of recombination crossref (matched: A general model for the evolution of recombination)
45 OK 10.1006/tpbi.1997.1301 Deleterious Mutations, Variable Epistatic Interactions, and the crossref (matched: Deleterious Mutations, Variable Epistatic Interactions, and the Evolut)
46 OK 10.1038/nrg761 Resolving the paradox of sex and recombination crossref (matched: Resolving the paradox of sex and recombination)
47 OK 10.1093/genetics/117.3.559 Selection, Generalized Transmission and the Evolution of Modifie crossref (matched: Selection, Generalized Transmission and the Evolution of Modifier Gene)
48 OK 10.1093/genetics/139.4.1805 The population genetics of speciation: the evolution of hybrid i crossref (matched: The population genetics of speciation: the evolution of hybrid incompa)
49 OK 10.1111/j.0014-3820.2001.tb00628.x THE EVOLUTION OF POSTZYGOTIC ISOLATION: ACCUMULATING DOBZHANSKY- crossref (matched: THE EVOLUTION OF POSTZYGOTIC ISOLATION: ACCUMULATING DOBZHANSKY-MULLER)
50 OK 10.48550/arXiv.2209.04836 Git Re-Basin: Merging Models modulo Permutation Symmetries in manuscript
51 OK 10.48550/arXiv.2606.23607 Scaling Linear Mode Connectivity and Merging to Billion Paramete in manuscript
52 OK 10.48550/arXiv.2410.12766 The Non-Local Model Merging Problem: Permutation Symmetries and in manuscript
53 OK 10.48550/arXiv.2607.11997 Are we Merging the Right Models? Impact of Expert Training Durat in manuscript
54 OK 10.48550/arXiv.2601.22285 Demystifying Mergeability: Interpretable Properties to Predict M in manuscript
55 OK 10.48550/arXiv.2205.12393 Fine-tuned Language Models are Continual Learners in manuscript
56 OK 10.48550/arXiv.2403.08763 Simple and Scalable Strategies to Continually Pre-train Large La in manuscript
57 OK 10.1080/09540099550039318 Catastrophic Forgetting, Rehearsal and Pseudorehearsal crossref (matched: Catastrophic Forgetting, Rehearsal and Pseudorehearsal)
58 OK 10.48550/arXiv.1705.08690 Continual Learning with Deep Generative Replay in manuscript
59 OK 10.48550/arXiv.2406.07515 Beyond Model Collapse: Scaling Up with Synthesized Data Requires in manuscript
60 OK 10.48550/arXiv.1606.04671 Progressive Neural Networks in manuscript
61 OK 10.48550/arXiv.2405.09673 LoRA Learns Less and Forgets Less in manuscript
62 OK 10.1037/0033-295x.102.3.419 Why there are complementary learning systems in the hippocampus crossref (matched: Why there are complementary learning systems in the hippocampus and ne)
63 OK 10.1016/j.tics.2016.05.004 What Learning Systems do Intelligent Agents Need? Complementary crossref (matched: What Learning Systems do Intelligent Agents Need? Complementary Learni)
64 OK 10.48550/arXiv.1805.06370 Progress &amp; Compress: A scalable framework for continual lear verified by hand
65 OK 10.48550/arXiv.2212.04089 Editing Models with Task Arithmetic in manuscript
66 OK 10.48550/arXiv.2407.06322 MagMax: Leveraging Model Merging for Seamless Continual Learning in manuscript
67 OK 10.48550/arXiv.2407.08699 Mitigating Catastrophic Forgetting in Language Transfer via Mode in manuscript
68 OK 10.48550/arXiv.2412.06712 How to Merge Your Multimodal Models Over Time? in manuscript
69 OK 10.48550/arXiv.1812.05159 An Empirical Study of Example Forgetting during Deep Neural Netw in manuscript
70 OK 10.48550/arXiv.2211.08411 Large Language Models Struggle to Learn Long-Tail Knowledge in manuscript
71 OK 10.48550/arXiv.2210.00266 Long-Tailed Class Incremental Learning in manuscript
72 OK 10.48550/arXiv.2309.10105 Understanding Catastrophic Forgetting in Language Models via Imp in manuscript
73 OK 10.48550/arXiv.2311.03099 Language Models are Super Mario: Absorbing Abilities from Homolo in manuscript
74 OK 10.48550/arXiv.2203.05482 Model soups: averaging weights of multiple fine-tuned models imp in manuscript
75 OK 10.48550/arXiv.2603.09463 An Empirical Study and Theoretical Explanation on Task-Level Mod in manuscript
76 OK 10.48550/arXiv.2506.14126 From Memorization to Parameter Interference: How Overtraining Ex in manuscript
77 OK 10.1145/2934662 Sex as an algorithm crossref (matched: Sex as an algorithm)
78 OK 10.48550/arXiv.2311.09807 The Curious Decline of Linguistic Diversity: Training Language M in manuscript
79 OK 10.48550/arXiv.2309.05196 Does Writing with Language Models Reduce Content Diversity? in manuscript
80 OK 10.1126/sciadv.adn5290 Generative AI enhances individual creativity but reduces the col crossref (matched: Generative AI enhances individual creativity but reduces the collectiv)

View file

@ -1,80 +0,0 @@
1 OK 10.48550/arXiv.2508.06811 Anatomy of a Machine Learning Ecosystem: 2 Million Models on Hug in manuscript
2 OK 10.48550/arXiv.2405.18432 Unsupervised Model Tree Heritage Recovery in manuscript
3 OK 10.48550/arXiv.2402.00699 PeaTMOSS: A Dataset and Initial Analysis of Pre-Trained Models i in manuscript
4 OK 10.48550/arXiv.2306.01708 TIES-Merging: Resolving Interference When Merging Models in manuscript
5 OK 10.1038/s42256-024-00975-8 Evolutionary optimization of model merging recipes crossref (matched: Evolutionary optimization of model merging recipes)
6 OK 10.48550/arXiv.2403.13257 Arcee's MergeKit: A Toolkit for Merging Large Language Models in manuscript
7 OK 10.48550/arXiv.2408.07666 Model Merging in LLMs, MLLMs, and Beyond: Methods, Theories, App in manuscript
8 OK 10.48550/arXiv.2503.01155 Nature-Inspired Population-Based Evolution of Large Language Mod in manuscript
9 OK 10.48550/arXiv.2508.16204 Competition and Attraction Improve Model Fusion in manuscript
10 OK 10.48550/arXiv.2501.05707 Multiagent Finetuning: Self Improvement with Diverse Reasoning C in manuscript
11 OK 10.48550/arXiv.2406.11704 Nemotron-4 340B Technical Report in manuscript
12 OK 10.48550/arXiv.2412.08905 Phi-4 Technical Report in manuscript
13 OK 10.48550/arXiv.2212.10560 Self-Instruct: Aligning Language Models with Self-Generated Inst in manuscript
14 OK 10.48550/arXiv.2401.05749 A Shocking Amount of the Web is Machine Translated: Insights fro in manuscript
15 OK 10.48550/arXiv.2403.07183 Monitoring AI-Modified Content at Scale: A Case Study on the Imp in manuscript
16 OK 10.48550/arXiv.2211.04325 Will we run out of data? Limits of LLM scaling based on human-ge in manuscript
17 OK 10.1038/s41562-023-01742-2 Machine culture verified by hand
18 OK 10.48550/arXiv.2304.03442 Generative Agents: Interactive Simulacra of Human Behavior in manuscript
19 OK 10.48550/arXiv.2402.01680 Large Language Model based Multi-Agents: A Survey of Progress an in manuscript
20 OK 10.48550/arXiv.2509.10147 Virtual Agent Economies in manuscript
21 OK 10.1038/s41586-024-07566-y AI models collapse when trained on recursively generated data crossref (matched: AI models collapse when trained on recursively generated data)
22 OK 10.1371/journal.pcbi.1002510 Structural Drift: The Population Dynamics of Sequential Learning crossref (matched: Structural Drift: The Population Dynamics of Sequential Learning)
23 OK 10.48550/arXiv.2604.08554 Drift and selection in LLM text ecosystems in manuscript
24 OK 10.48550/arXiv.2509.20101 First-Extinction Law for Resampling Processes in manuscript
25 OK 10.48550/arXiv.2407.17493 Model Collapse in the Self-Consuming Chain of Diffusion Finetuni in manuscript
26 OK 10.1016/s0079-7421(08)60536-8 Catastrophic Interference in Connectionist Networks: The Sequent crossref (matched: Catastrophic Interference in Connectionist Networks: The Sequential Le)
27 OK 10.1016/s1364-6613(99)01294-2 Catastrophic forgetting in connectionist networks crossref (matched: Catastrophic forgetting in connectionist networks)
28 OK 10.1016/0027-5107(64)90047-8 The relation of recombination to mutational advance crossref (matched: The relation of recombination to mutational advance)
29 OK 10.48550/arXiv.2510.16657 Escaping Model Collapse via Synthetic Data Verification: Near-te in manuscript
30 OK 10.48550/arXiv.2404.01413 Is Model Collapse Inevitable? Breaking the Curse of Recursion by in manuscript
31 OK 10.1093/genetics/16.2.97 EVOLUTION IN MENDELIAN POPULATIONS crossref (matched: EVOLUTION IN MENDELIAN POPULATIONS)
32 OK 10.1046/j.1523-1739.1996.10061509.x The OneMigrantperGeneration Rule in Conservation and Manageme crossref (matched: The OneMigrantperGeneration Rule in Conservation and Management)
33 HAND (pre-DOI source) [Review of] The Origin of Species written by hand
34 OK 10.48550/arXiv.2411.02207 Collective Model Intelligence Requires Compatible Specialization in manuscript
35 HAND (pre-DOI source) The Genetical Theory of Natural Selection written by hand
36 OK 10.1086/280418 Some Genetic Aspects of Sex crossref (matched: Some Genetic Aspects of Sex)
37 OK 10.48550/arXiv.2106.09685 LoRA: Low-Rank Adaptation of Large Language Models in manuscript
38 OK 10.1016/s0022-5193(87)80029-2 Towards a general theory of adaptive walks on rugged landscapes crossref (matched: Towards a general theory of adaptive walks on rugged landscapes)
39 HAND (pre-DOI source) Coadaptation and outbreeding depression written by hand
40 OK 10.1162/evco_a_00025 Abandoning Objectives: Evolution Through the Search for Novelty crossref (matched: Abandoning Objectives: Evolution Through the Search for Novelty Alone)
41 OK 10.48550/arXiv.2503.05683 WikiBigEdit: Understanding the Limits of Lifelong Knowledge Edit in manuscript
42 OK 10.48550/arXiv.2502.04390 In Praise of Stubbornness: An Empirical Case for Cognitive-Disso in manuscript
43 OK 10.48550/arXiv.2607.09202 Interference and Retention in Continual Learning in manuscript
44 OK 10.1017/s0016672300033140 A general model for the evolution of recombination crossref (matched: A general model for the evolution of recombination)
45 OK 10.1006/tpbi.1997.1301 Deleterious Mutations, Variable Epistatic Interactions, and the crossref (matched: Deleterious Mutations, Variable Epistatic Interactions, and the Evolut)
46 OK 10.1038/nrg761 Resolving the paradox of sex and recombination crossref (matched: Resolving the paradox of sex and recombination)
47 OK 10.1093/genetics/117.3.559 Selection, Generalized Transmission and the Evolution of Modifie crossref (matched: Selection, Generalized Transmission and the Evolution of Modifier Gene)
48 OK 10.1093/genetics/139.4.1805 The population genetics of speciation: the evolution of hybrid i crossref (matched: The population genetics of speciation: the evolution of hybrid incompa)
49 OK 10.1111/j.0014-3820.2001.tb00628.x THE EVOLUTION OF POSTZYGOTIC ISOLATION: ACCUMULATING DOBZHANSKY- crossref (matched: THE EVOLUTION OF POSTZYGOTIC ISOLATION: ACCUMULATING DOBZHANSKY-MULLER)
50 OK 10.48550/arXiv.2209.04836 Git Re-Basin: Merging Models modulo Permutation Symmetries in manuscript
51 OK 10.48550/arXiv.2606.23607 Scaling Linear Mode Connectivity and Merging to Billion Paramete in manuscript
52 OK 10.48550/arXiv.2410.12766 The Non-Local Model Merging Problem: Permutation Symmetries and in manuscript
53 OK 10.48550/arXiv.2607.11997 Are we Merging the Right Models? Impact of Expert Training Durat in manuscript
54 OK 10.48550/arXiv.2601.22285 Demystifying Mergeability: Interpretable Properties to Predict M in manuscript
55 OK 10.48550/arXiv.2205.12393 Fine-tuned Language Models are Continual Learners in manuscript
56 OK 10.48550/arXiv.2403.08763 Simple and Scalable Strategies to Continually Pre-train Large La in manuscript
57 OK 10.1080/09540099550039318 Catastrophic Forgetting, Rehearsal and Pseudorehearsal crossref (matched: Catastrophic Forgetting, Rehearsal and Pseudorehearsal)
58 OK 10.48550/arXiv.1705.08690 Continual Learning with Deep Generative Replay in manuscript
59 OK 10.48550/arXiv.2406.07515 Beyond Model Collapse: Scaling Up with Synthesized Data Requires in manuscript
60 OK 10.48550/arXiv.1606.04671 Progressive Neural Networks in manuscript
61 OK 10.48550/arXiv.2405.09673 LoRA Learns Less and Forgets Less in manuscript
62 OK 10.1037/0033-295x.102.3.419 Why there are complementary learning systems in the hippocampus crossref (matched: Why there are complementary learning systems in the hippocampus and ne)
63 OK 10.1016/j.tics.2016.05.004 What Learning Systems do Intelligent Agents Need? Complementary crossref (matched: What Learning Systems do Intelligent Agents Need? Complementary Learni)
64 OK 10.48550/arXiv.1805.06370 Progress &amp; Compress: A scalable framework for continual lear verified by hand
65 OK 10.48550/arXiv.2212.04089 Editing Models with Task Arithmetic in manuscript
66 OK 10.48550/arXiv.2407.06322 MagMax: Leveraging Model Merging for Seamless Continual Learning in manuscript
67 OK 10.48550/arXiv.2407.08699 Mitigating Catastrophic Forgetting in Language Transfer via Mode in manuscript
68 OK 10.48550/arXiv.2412.06712 How to Merge Your Multimodal Models Over Time? in manuscript
69 OK 10.48550/arXiv.1812.05159 An Empirical Study of Example Forgetting during Deep Neural Netw in manuscript
70 OK 10.48550/arXiv.2211.08411 Large Language Models Struggle to Learn Long-Tail Knowledge in manuscript
71 OK 10.48550/arXiv.2210.00266 Long-Tailed Class Incremental Learning in manuscript
72 OK 10.48550/arXiv.2309.10105 Understanding Catastrophic Forgetting in Language Models via Imp in manuscript
73 OK 10.48550/arXiv.2311.03099 Language Models are Super Mario: Absorbing Abilities from Homolo in manuscript
74 OK 10.48550/arXiv.2203.05482 Model soups: averaging weights of multiple fine-tuned models imp in manuscript
75 OK 10.48550/arXiv.2603.09463 An Empirical Study and Theoretical Explanation on Task-Level Mod in manuscript
76 OK 10.48550/arXiv.2506.14126 From Memorization to Parameter Interference: How Overtraining Ex in manuscript
77 OK 10.1145/2934662 Sex as an algorithm crossref (matched: Sex as an algorithm)
78 OK 10.48550/arXiv.2311.09807 The Curious Decline of Linguistic Diversity: Training Language M in manuscript
79 OK 10.48550/arXiv.2309.05196 Does Writing with Language Models Reduce Content Diversity? in manuscript
80 OK 10.1126/sciadv.adn5290 Generative AI enhances individual creativity but reduces the col crossref (matched: Generative AI enhances individual creativity but reduces the collectiv)
1 1 OK 10.48550/arXiv.2508.06811 Anatomy of a Machine Learning Ecosystem: 2 Million Models on Hug in manuscript
2 2 OK 10.48550/arXiv.2405.18432 Unsupervised Model Tree Heritage Recovery in manuscript
3 3 OK 10.48550/arXiv.2402.00699 PeaTMOSS: A Dataset and Initial Analysis of Pre-Trained Models i in manuscript
4 4 OK 10.48550/arXiv.2306.01708 TIES-Merging: Resolving Interference When Merging Models in manuscript
5 5 OK 10.1038/s42256-024-00975-8 Evolutionary optimization of model merging recipes crossref (matched: Evolutionary optimization of model merging recipes)
6 6 OK 10.48550/arXiv.2403.13257 Arcee's MergeKit: A Toolkit for Merging Large Language Models in manuscript
7 7 OK 10.48550/arXiv.2408.07666 Model Merging in LLMs, MLLMs, and Beyond: Methods, Theories, App in manuscript
8 8 OK 10.48550/arXiv.2503.01155 Nature-Inspired Population-Based Evolution of Large Language Mod in manuscript
9 9 OK 10.48550/arXiv.2508.16204 Competition and Attraction Improve Model Fusion in manuscript
10 10 OK 10.48550/arXiv.2501.05707 Multiagent Finetuning: Self Improvement with Diverse Reasoning C in manuscript
11 11 OK 10.48550/arXiv.2406.11704 Nemotron-4 340B Technical Report in manuscript
12 12 OK 10.48550/arXiv.2412.08905 Phi-4 Technical Report in manuscript
13 13 OK 10.48550/arXiv.2212.10560 Self-Instruct: Aligning Language Models with Self-Generated Inst in manuscript
14 14 OK 10.48550/arXiv.2401.05749 A Shocking Amount of the Web is Machine Translated: Insights fro in manuscript
15 15 OK 10.48550/arXiv.2403.07183 Monitoring AI-Modified Content at Scale: A Case Study on the Imp in manuscript
16 16 OK 10.48550/arXiv.2211.04325 Will we run out of data? Limits of LLM scaling based on human-ge in manuscript
17 17 OK 10.1038/s41562-023-01742-2 Machine culture verified by hand
18 18 OK 10.48550/arXiv.2304.03442 Generative Agents: Interactive Simulacra of Human Behavior in manuscript
19 19 OK 10.48550/arXiv.2402.01680 Large Language Model based Multi-Agents: A Survey of Progress an in manuscript
20 20 OK 10.48550/arXiv.2509.10147 Virtual Agent Economies in manuscript
21 21 OK 10.1038/s41586-024-07566-y AI models collapse when trained on recursively generated data crossref (matched: AI models collapse when trained on recursively generated data)
22 22 OK 10.1371/journal.pcbi.1002510 Structural Drift: The Population Dynamics of Sequential Learning crossref (matched: Structural Drift: The Population Dynamics of Sequential Learning)
23 23 OK 10.48550/arXiv.2604.08554 Drift and selection in LLM text ecosystems in manuscript
24 24 OK 10.48550/arXiv.2509.20101 First-Extinction Law for Resampling Processes in manuscript
25 25 OK 10.48550/arXiv.2407.17493 Model Collapse in the Self-Consuming Chain of Diffusion Finetuni in manuscript
26 26 OK 10.1016/s0079-7421(08)60536-8 Catastrophic Interference in Connectionist Networks: The Sequent crossref (matched: Catastrophic Interference in Connectionist Networks: The Sequential Le)
27 27 OK 10.1016/s1364-6613(99)01294-2 Catastrophic forgetting in connectionist networks crossref (matched: Catastrophic forgetting in connectionist networks)
28 28 OK 10.1016/0027-5107(64)90047-8 The relation of recombination to mutational advance crossref (matched: The relation of recombination to mutational advance)
29 29 OK 10.48550/arXiv.2510.16657 Escaping Model Collapse via Synthetic Data Verification: Near-te in manuscript
30 30 OK 10.48550/arXiv.2404.01413 Is Model Collapse Inevitable? Breaking the Curse of Recursion by in manuscript
31 31 OK 10.1093/genetics/16.2.97 EVOLUTION IN MENDELIAN POPULATIONS crossref (matched: EVOLUTION IN MENDELIAN POPULATIONS)
32 32 OK 10.1046/j.1523-1739.1996.10061509.x The One‐Migrant‐per‐Generation Rule in Conservation and Manageme crossref (matched: The One‐Migrant‐per‐Generation Rule in Conservation and Management)
33 33 HAND (pre-DOI source) [Review of] The Origin of Species written by hand
34 34 OK 10.48550/arXiv.2411.02207 Collective Model Intelligence Requires Compatible Specialization in manuscript
35 35 HAND (pre-DOI source) The Genetical Theory of Natural Selection written by hand
36 36 OK 10.1086/280418 Some Genetic Aspects of Sex crossref (matched: Some Genetic Aspects of Sex)
37 37 OK 10.48550/arXiv.2106.09685 LoRA: Low-Rank Adaptation of Large Language Models in manuscript
38 38 OK 10.1016/s0022-5193(87)80029-2 Towards a general theory of adaptive walks on rugged landscapes crossref (matched: Towards a general theory of adaptive walks on rugged landscapes)
39 39 HAND (pre-DOI source) Coadaptation and outbreeding depression written by hand
40 40 OK 10.1162/evco_a_00025 Abandoning Objectives: Evolution Through the Search for Novelty crossref (matched: Abandoning Objectives: Evolution Through the Search for Novelty Alone)
41 41 OK 10.48550/arXiv.2503.05683 WikiBigEdit: Understanding the Limits of Lifelong Knowledge Edit in manuscript
42 42 OK 10.48550/arXiv.2502.04390 In Praise of Stubbornness: An Empirical Case for Cognitive-Disso in manuscript
43 43 OK 10.48550/arXiv.2607.09202 Interference and Retention in Continual Learning in manuscript
44 44 OK 10.1017/s0016672300033140 A general model for the evolution of recombination crossref (matched: A general model for the evolution of recombination)
45 45 OK 10.1006/tpbi.1997.1301 Deleterious Mutations, Variable Epistatic Interactions, and the crossref (matched: Deleterious Mutations, Variable Epistatic Interactions, and the Evolut)
46 46 OK 10.1038/nrg761 Resolving the paradox of sex and recombination crossref (matched: Resolving the paradox of sex and recombination)
47 47 OK 10.1093/genetics/117.3.559 Selection, Generalized Transmission and the Evolution of Modifie crossref (matched: Selection, Generalized Transmission and the Evolution of Modifier Gene)
48 48 OK 10.1093/genetics/139.4.1805 The population genetics of speciation: the evolution of hybrid i crossref (matched: The population genetics of speciation: the evolution of hybrid incompa)
49 49 OK 10.1111/j.0014-3820.2001.tb00628.x THE EVOLUTION OF POSTZYGOTIC ISOLATION: ACCUMULATING DOBZHANSKY- crossref (matched: THE EVOLUTION OF POSTZYGOTIC ISOLATION: ACCUMULATING DOBZHANSKY-MULLER)
50 50 OK 10.48550/arXiv.2209.04836 Git Re-Basin: Merging Models modulo Permutation Symmetries in manuscript
51 51 OK 10.48550/arXiv.2606.23607 Scaling Linear Mode Connectivity and Merging to Billion Paramete in manuscript
52 52 OK 10.48550/arXiv.2410.12766 The Non-Local Model Merging Problem: Permutation Symmetries and in manuscript
53 53 OK 10.48550/arXiv.2607.11997 Are we Merging the Right Models? Impact of Expert Training Durat in manuscript
54 54 OK 10.48550/arXiv.2601.22285 Demystifying Mergeability: Interpretable Properties to Predict M in manuscript
55 55 OK 10.48550/arXiv.2205.12393 Fine-tuned Language Models are Continual Learners in manuscript
56 56 OK 10.48550/arXiv.2403.08763 Simple and Scalable Strategies to Continually Pre-train Large La in manuscript
57 57 OK 10.1080/09540099550039318 Catastrophic Forgetting, Rehearsal and Pseudorehearsal crossref (matched: Catastrophic Forgetting, Rehearsal and Pseudorehearsal)
58 58 OK 10.48550/arXiv.1705.08690 Continual Learning with Deep Generative Replay in manuscript
59 59 OK 10.48550/arXiv.2406.07515 Beyond Model Collapse: Scaling Up with Synthesized Data Requires in manuscript
60 60 OK 10.48550/arXiv.1606.04671 Progressive Neural Networks in manuscript
61 61 OK 10.48550/arXiv.2405.09673 LoRA Learns Less and Forgets Less in manuscript
62 62 OK 10.1037/0033-295x.102.3.419 Why there are complementary learning systems in the hippocampus crossref (matched: Why there are complementary learning systems in the hippocampus and ne)
63 63 OK 10.1016/j.tics.2016.05.004 What Learning Systems do Intelligent Agents Need? Complementary crossref (matched: What Learning Systems do Intelligent Agents Need? Complementary Learni)
64 64 OK 10.48550/arXiv.1805.06370 Progress &amp; Compress: A scalable framework for continual lear verified by hand
65 65 OK 10.48550/arXiv.2212.04089 Editing Models with Task Arithmetic in manuscript
66 66 OK 10.48550/arXiv.2407.06322 MagMax: Leveraging Model Merging for Seamless Continual Learning in manuscript
67 67 OK 10.48550/arXiv.2407.08699 Mitigating Catastrophic Forgetting in Language Transfer via Mode in manuscript
68 68 OK 10.48550/arXiv.2412.06712 How to Merge Your Multimodal Models Over Time? in manuscript
69 69 OK 10.48550/arXiv.1812.05159 An Empirical Study of Example Forgetting during Deep Neural Netw in manuscript
70 70 OK 10.48550/arXiv.2211.08411 Large Language Models Struggle to Learn Long-Tail Knowledge in manuscript
71 71 OK 10.48550/arXiv.2210.00266 Long-Tailed Class Incremental Learning in manuscript
72 72 OK 10.48550/arXiv.2309.10105 Understanding Catastrophic Forgetting in Language Models via Imp in manuscript
73 73 OK 10.48550/arXiv.2311.03099 Language Models are Super Mario: Absorbing Abilities from Homolo in manuscript
74 74 OK 10.48550/arXiv.2203.05482 Model soups: averaging weights of multiple fine-tuned models imp in manuscript
75 75 OK 10.48550/arXiv.2603.09463 An Empirical Study and Theoretical Explanation on Task-Level Mod in manuscript
76 76 OK 10.48550/arXiv.2506.14126 From Memorization to Parameter Interference: How Overtraining Ex in manuscript
77 77 OK 10.1145/2934662 Sex as an algorithm crossref (matched: Sex as an algorithm)
78 78 OK 10.48550/arXiv.2311.09807 The Curious Decline of Linguistic Diversity: Training Language M in manuscript
79 79 OK 10.48550/arXiv.2309.05196 Does Writing with Language Models Reduce Content Diversity? in manuscript
80 80 OK 10.1126/sciadv.adn5290 Generative AI enhances individual creativity but reduces the col crossref (matched: Generative AI enhances individual creativity but reduces the collectiv)

View file

@ -1,131 +0,0 @@
"""Renumber the manuscript's references to first-appearance order (PNAS style).
Reads paper/manuscript/main.md, finds every parenthesised citation group in the text above
"## References", derives the order in which references first appear, and rewrites the citation
groups in main.md, si.md, and the figure captions in build.py, then reorders the reference list.
Citation groups are parentheses containing only reference numbers, commas, en-dash ranges, an optional
"cf. " prefix, or a prose prefix ending in a semicolon ("...; 11, 12"). Four-digit numbers (years)
never match, and any number above the list length is reported and left alone.
Usage: python paper/manuscript/renumber_refs.py # dry run: mapping + per-file counts
python paper/manuscript/renumber_refs.py --apply # rewrite the three files in place
"""
from __future__ import annotations
import re
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parents[2]
MAIN, SI, BUILD = (ROOT / "paper" / "manuscript" / n for n in ("main.md", "si.md", "build.py"))
REF_HEADER = "## References"
CIT = re.compile(
r"\((?P<pre>[^()]*?;\s*)?(?P<cf>cf\.\s*)?"
r"(?P<nums>\d{1,3}(?:\s*[-]\s*\d{1,3})?(?:,\s*\d{1,3}(?:\s*[-]\s*\d{1,3})?)*)\)"
)
REF_LINE = re.compile(r"^(\d+)\. (.*)$")
def expand(nums: str) -> list[int]:
out: list[int] = []
for part in re.split(r",\s*", nums):
if re.search(r"[-]", part):
a, b = (int(x) for x in re.split(r"\s*[-]\s*", part))
out.extend(range(a, b + 1))
else:
out.append(int(part))
return out
def compress(nums: list[int]) -> str:
"""Ascending, with runs of three or more collapsed to an en-dash range."""
nums = sorted(set(nums))
runs: list[list[int]] = []
for n in nums:
if runs and n == runs[-1][-1] + 1:
runs[-1].append(n)
else:
runs.append([n])
return ", ".join(f"{r[0]}{r[-1]}" if len(r) >= 3 else ", ".join(map(str, r)) for r in runs)
def split_main(text: str) -> tuple[str, list[tuple[int, str]]]:
body, _, refs = text.partition(REF_HEADER)
entries = [(int(m.group(1)), m.group(2)) for line in refs.splitlines() if (m := REF_LINE.match(line))]
return body, entries
def first_appearance(body: str, n_refs: int) -> list[int]:
order: list[int] = []
for m in CIT.finditer(body):
for n in expand(m.group("nums")):
if n <= n_refs and n not in order:
order.append(n)
return order
def rewrite(text: str, mapping: dict[int, int], n_refs: int, label: str) -> tuple[str, int, list[str]]:
count, suspicious = 0, []
def sub(m: re.Match) -> str:
nonlocal count
nums = expand(m.group("nums"))
if any(n > n_refs or n < 1 for n in nums):
suspicious.append(m.group(0))
return m.group(0)
count += 1
return f"({m.group('pre') or ''}{m.group('cf') or ''}{compress([mapping[n] for n in nums])})"
return CIT.sub(sub, text), count, suspicious
def main(apply: bool) -> int:
main_text = MAIN.read_text()
body, entries = split_main(main_text)
n_refs = len(entries)
assert [n for n, _ in entries] == list(range(1, n_refs + 1)), "reference list is not 1..N"
order = first_appearance(body, n_refs)
orphans = sorted(set(range(1, n_refs + 1)) - set(order))
if orphans:
print(f"ERROR: never cited in main text: {orphans}")
return 1
mapping = {old: new for new, old in enumerate(order, start=1)}
changed = {o: n for o, n in mapping.items() if o != n}
print(f"{n_refs} references; {len(changed)} renumbered" + (":" if changed else "."))
for o in sorted(changed):
print(f" {o:3d} -> {mapping[o]:3d} {entries[o - 1][1][:70]}")
outputs: dict[Path, str] = {}
new_body, c, sus = rewrite(body, mapping, n_refs, "main")
print(f"main.md: {c} citation groups" + (f"; left alone: {sus}" if sus else ""))
by_new = sorted(entries, key=lambda e: mapping[e[0]])
new_refs = "\n".join(f"{mapping[o]}. {t}" for o, t in by_new)
outputs[MAIN] = f"{new_body}{REF_HEADER}\n\n{new_refs}\n"
text, c, sus = rewrite(SI.read_text(), mapping, n_refs, SI.name)
print(f"{SI.name}: {c} citation groups" + (f"; left alone: {sus}" if sus else ""))
outputs[SI] = text
# build.py is Python: only its FIGURES caption block may carry citations, so rewrite that slice
# alone — tuples like (0, 1) elsewhere in the code would otherwise look like citations.
btext = BUILD.read_text()
head = re.search(r"^FIGURES\b[^\n]*\{\s*$", btext, re.M)
if head is None:
print(f"{BUILD.name}: no FIGURES block found; skipped")
else:
start = head.start()
end = btext.index("\n}\n", start) + 3
block, c, sus = rewrite(btext[start:end], mapping, n_refs, BUILD.name)
print(f"{BUILD.name} captions: {c} citation groups" + (f"; left alone: {sus}" if sus else ""))
outputs[BUILD] = btext[:start] + block + btext[end:]
if apply:
for path, text in outputs.items():
path.write_text(text)
print("applied.")
else:
print("dry run — pass --apply to write.")
return 0
if __name__ == "__main__":
sys.exit(main("--apply" in sys.argv))

Binary file not shown.

View file

@ -1,146 +0,0 @@
# Response to the second review
*This response accompanies a further revision. Every number below is reproduced by a committed script
(`figures/stats_llm_epistasis.py`) from committed artifacts; the revised documents are
`results/llm_epistasis/README.md` (rewritten around your four analyses), the manuscript, and
`paper/si-notes.md`.*
---
## 0. A correction first (your §7)
You are right, and we apologise for the bookkeeping error: the draft you reviewed **already
contained** the full-symmetry alignment experiment, the conflict residual, and the
compatible-specialisation null. Our previous letter's "new since the review" conflated three things
that we now state separately: **new results** (the LLM-tier speciation runs, the multi-seed
replication, and the controlled predictive test with its control axis), **new controls and analyses**
(the compatible-overlap axis; the robust statistics in this letter), and **revised interpretation**
(everything about E13b/c, which was experimental content you had already seen and whose *presentation*
we changed). The experimental delta attributable to the review process is the first and second
categories only.
## 1. Your two-conclusion distinction — adopted as the evidential boundary
We accept the boundary exactly as you drew it:
- **Demonstrated:** a small-model, controlled predictive test in which pre-merge functional
disagreement predicted merge penalties where the selected weight-space measures did not.
- **Not demonstrated:** that an epistasis-specific predictor adds value beyond ordinary functional
disagreement, or that the prediction improves operator choice.
The experiment is now titled a **"controlled predictive test"** everywhere ("decisive experiment
delivered" is gone), and the manuscript's §1 ladder describes the prediction rung in your conditional
formulation, with its four boundary clauses stated in place: constructed grid, small scale, refinement
not superior, operator choice open. The measure itself is renamed **"confidence-weighted functional
conflict — a proposed proxy for merge-relevant interactions"**; we accept that bilateral confident
contradiction measures incompatible endpoint behaviour, not non-additive interaction in the strict
biological sense, and the paper no longer calls any measured quantity "epistasis." Your distinction —
*the framework motivated the measurement and controls* vs *their success validates the specifically
population-genetic mechanism* — is adopted verbatim; we claim the former.
## 2. The four analyses — run
**(1) Direct predictor comparison.** Condition-clustered bootstrap (13 clusters, B = 4000), 95% CIs
for each predictor's ρ against the pre-registered primary outcome:
| predictor | ρ | clustered 95% CI |
|---|---|---|
| raw functional disagreement | +0.460 | [+0.04, +0.69] |
| confidence-weighted functional conflict | +0.446 | [+0.02, +0.68] |
| gradient alignment | 0.347 | [0.59, 0.06] |
| delta L2 | +0.165 | [0.27, +0.58] |
| delta cosine | +0.030 | [0.46, +0.51] |
| cross-family accuracy | 0.005 | [0.29, +0.31] |
**Paired contrasts are not individually significant** (e.g. |ρ(dis_raw)| |ρ(delta_cos)| = +0.23,
CI [0.23, +0.59]). Held-out prediction (leave-one-condition-out linear fits): functional measures
replicate (dis_raw ρ = +0.396, p = 0.013; conf-weighted +0.352, p = 0.028); geometry ≈ 0; the
performance baseline is unstable out-of-sample (0.435). So the supported statement — now the
conclusion in the README, the figure title, and the manuscript — is yours: *across this controlled
grid, functional disagreement showed a detectable, held-out-robust association with merge penalty;
LoRA-delta cosine and L2 showed no statistically detectable association; gradient alignment carried
intermediate signal (its CI excludes zero), so this is not a clean functional-versus-all-geometric
divide; head-to-head predictor differences are not individually significant; only these baselines were
tested.* "Weight divergence does not predict merge failure" has been removed as over-broad.
**(2) Sample structure.** 39 rows = **13 conditions × 3 seeds**. Parents are retrained per
condition × seed, but children share task-data seeds across conditions within a seed (e.g. the
duration-3 parent and the conflict-0 parent are trained on essentially the same data), so rows are not
independent — hence clustering by condition in all uncertainty estimates, disclosed in the README.
**(3) Between- vs within-axis.** You were right that the pooled correlation is substantially axis
discrimination (mean penalty: conflict 0.061 vs compat 0.005 / duration 0.011). Within the conflict
axis (n = 15): functional measures +0.59/+0.62 — but delta-L2 is +0.71 there, because *within that
axis* conflict fraction, added-data volume, and delta growth are collinear: **within-axis
identification is impossible by design**, and we now say so; the identification comes from the control
axes, where the same volumes and L2 ranges occur at ~zero penalty. Scatterplots are coloured by axis
in the figure. Predicting penalties for a held-out *conflict mechanism* (not just held-out conditions
of the same mechanism) is listed as the next test — we agree it is the more valuable one.
**(4) Outcome references.** Reported under all three: oracle parent potential (pre-registered
primary), best parent, and mean parent. The ordering is **sensitive to the reference** — under the
best-parent reference, delta-L2 correlates comparably to the functional measures (+0.48 vs +0.34).
Our reading, stated in the README rather than hidden: that reference inherits parent-strength trends
that track training volume, which geometry also tracks, coupling predictor and outcome through the
reference rather than through merge damage. On structural coupling between the disagreement predictor
and the oracle-potential outcome: probe and test sets are disjoint by construction, but we agree
definitional coupling through parental complementarity cannot be fully excluded, which is one more
reason all three references are now on the table.
## 3. Hypothesis labels
Your table is adopted essentially as written — the manuscript's claims-at-a-glance table now carries:
pre-merge disagreement predicts penalty (*empirical within the controlled grid*, with the boundary
clauses in the limits column); confidence weighting improves rank prediction (***not supported***);
functional beats all geometry (*not established — selected baselines only*); operator choice (*open*);
cliff, snowball, emergent DMIs (*hypotheses*, unchanged). We also added your snowball distinction
where the snowball is discussed: super-linear growth in incompatibility *count* does not by itself
entail a sharp *performance* cliff — that needs the count→effect-size→performance link, which the
analytic model supplies under its assumptions and any neural test must establish separately.
## 4. The two technical statements — corrected
**"Endpoints and chord are invariant."** Accepted; the ambiguity was ours. The SI proposition now
defines "chord" precisely as the α-linear interpolation **of the endpoint loss values** — the barrier
baseline, a function of endpoints only, which *is* invariant — and states explicitly that the
**weight-space interpolation path is generally not invariant** (that being precisely why alignment can
lower a barrier). It also now carries your second point: exact recovery of a permuted-and-rescaled
copy validates a special case and does not establish global optimality of the alignment for
independently trained networks — so the "removable" share is a lower bound and the "residual" an upper
bound, stated wherever the decomposition is used.
**"Frozen base pins the coordinate system, so failure is functional by construction."** Accepted, and
your replacement wording adopted verbatim across the module, configs, READMEs, and figure: *the shared
frozen base controls a major source of coordinate mismatch, allowing a cleaner test of
conflict-associated merging failure* — with the boundary stated (failures of delta-averaging can still
reflect nonlinear interaction, scaling, or capacity). "Conflict-associated," not "functional by
construction," throughout.
## 5. Chronology and reliability
**Chronology (now a section of the README).** Prospective: hypotheses, predictors, primary outcome and
falsifiers were in the config before the first grid ran. Adaptive: the compatible-overlap control was
added *after* geometry appeared to win, with its own pre-stated readings, run on the same seeds; no
existing rows were re-run or altered. Post hoc: the clustered-bootstrap/LOCO/multi-reference analyses
were added at your request after all data was collected. We agree this is transparent adaptive
experimentation, not wholly prospective confirmation, and it is labelled as such.
**Reliability.** The CI-width claim is withdrawn. The seed-level statement now reads: routing beat the
soup **in every seed** (3/3 paired, both metrics), directed selection beat the soup 3/3, and one seed
exhibited a catastrophic soup failure (0.071 overall, 0.000 worst-family) to which routing was immune
(0.262/0.225); seed-level sds (0.090 vs 0.023) are reported as an observation, with the explicit note
that three seeds do not support a variance estimate.
## 6. On your bottom line
We accept your formulation as the paper's claim for this line of work — it now closes the relevant
results section nearly verbatim: *we separated overlap, divergence, and conflict experimentally; in
this controlled setting, functional disagreement predicted merging damage when simple weight-distance
measures did not; the proposed epistasis refinement and the emergent-speciation mechanism remain
unconfirmed.* And we take the redirection about what would count next: not 0.5B → 7B alone, but
**generalisation to unfamiliar conflict structures** (a held-out conflict mechanism, and real rather
than constructed task pairs) and **a demonstrably better budget-matched merging decision**. Those two
now head the open-problems list, above the scale replication.
We would welcome a third pass if you have the appetite — particularly on whether the README's
conditional conclusion and the chronology section read at the right strength.

View file

@ -1,96 +0,0 @@
# Response to the third review (of the PNAS-format draft)
*All five priority fixes are made, plus the presentation items. The revised draft is
`paper/manuscript/main.md` (rebuilt PDF alongside); the long-form document and the results documentation
were corrected wherever they carried the same overstatements. Point-by-point:*
## 1. The averaging proposition (your §2) — you are right, and the text now proves what it claims
Your convexity argument is correct: conservation of expected mass does not establish that averaging
cannot help, because extinction is convex in mixed mass and averaging reduces its variance. Our result
is, exactly as you diagnosed, a **first-order cancellation in the rare-item regime**, and the main
text now states the actual proposition with its quantities and assumptions: K parents with independent
retention; child draws `n` samples from one random parent vs the parents' output-mean; expected mass
identical; and in the regime `n·p/K ≪ 1`, where per-item survival is first-order in sampled mass,
expected survival is identical too. Two boundaries follow in the same paragraph: outside that regime
averaging's variance reduction can *reduce* extinction relative to a random single parent (your
argument, credited to the review process); and the union operator's renormalisation (which itself
redistributes mass) and oracle requirement are stated. "Adding parents cannot help" is deleted here
and in every other document that carried it. We agree the interesting content is the consequence for
retention, not the elementary conservation of a mean — which is how the proposition is now framed.
## 2. Grounding (your §3) — threshold made operational, floor made probabilistic, rule de-categoricalised
- `g* ≈ 0.05` is now explicitly an **operational threshold**, with the text stating what our own
analysis always showed: the immigrationdrift equilibrium is *smooth* in the grounding fraction (no
phase transition in aggregate diversity). New wording: under the tested population size and Zipf
source, `g ≈ 0.05` retained ≥95% of equilibrium diversity, with dependence on sample size, source,
and retention target (SI).
- `m·p ≳ 1` is restated as what it is: `1 e^{m·p}` observation probability per batch (~63% at
`m·p = 1`), confidence-dependent, with retention vs stationary occupancy vs reintroduction
distinguished (immigration can restore an absent item).
- The design rule now reads in your form: under unstratified grounding rare capabilities are expensive
(targeted sampling changes the cost); recombination recovers rare capabilities *still retained
across complementary parents*.
## 3. Grounded inheritance vs grounded evaluation (your §4) — separated and named
The society section now opens with the definitional distinction: **grounded inheritance** (external
samples in the reproduction process — the data channel) vs **grounded evaluation** (true fitness vs
conformity in selection — the fitness channel), related but different operators, connected only in
that both couple the lineage to a non-drifting external signal. The section is retitled to your
formulation ("…make complementary contributions"), the ablation is described as separating failure
modes *under the tested conditions*, and general joint necessity is explicitly disclaimed (alternative
mutation/restart/archive/selection schemes noted). Table 1's corresponding row now says
"complementary… in the tested society"; the same fix is propagated to the long-form document.
## 4. The alignment contradiction (your §5) — deleted, both statements reconciled
"This cannot be an alignment failure, because the same aligner succeeded on the control" is removed
everywhere (manuscript, long-form document, results documentation), replaced by your formulation: the
tested alignment removes the same-task barrier but leaves the conflict-associated barrier largely
unchanged — supporting a functional-conflict interpretation without proving optimal alignment. The
abstract now says "remaining after permutation-and-rescaling alignment" (not "surviving the full
symmetry group"), and the Methods note that the group is the alignment's *search space*, with control
recovery not establishing global optimality. The discussion's "expect specialisation alone to be
merge-safe" is replaced by the supported lesson: **do not treat divergence or specialisation alone as
evidence of incompatibility.**
## 5. Headline vs detail (your §6) — matched, and the seed-dependence analysed
The significance statement now ends with your suggested sentence (a controlled small-model test…
motivating further comparison). On the clustering point: you are right that condition-clustering does
not capture cross-condition dependence through shared task-data seeds. We added the sensitivity you
asked for (committed to the statistics script): **per-seed correlations** — each seed alone, n = 13
conditions — are stable for the functional measures (+0.37 to +0.53 in every individual seed) and ≈0
for geometry in every seed; leave-one-seed-out ranges are [+0.38, +0.56] (functional) vs
[0.04, +0.28] (geometry). One informative surprise: gradient alignment is *seed-unstable*
(0.11 to 0.55), which the manuscript now reports as its own caveat. The text also states plainly
that with three seeds, uncertainty about seed generalisation remains substantial.
## 6. Presentation (your §7) — done
Meta-language removed ("the honest statement", "sharpest honesty", "earn their place by tempering",
"honest deviations" — all gone; results are stated, not described as disclosures). "Exact" is now
reserved for closed-form mathematics — NK/simulation results are labelled "analytic model" in Table 1
and the text. The headroom relationship is stated qualitatively with "a quantitative form is
untested". "Directed sex with no biological analogue" is replaced by your phrasing (the shorthand kept,
defined as engineered recombination with flexible parent choice and pre-deployment screening).
Muller's ratchet is now a *consequence-level* correspondence, with the text stating that irreversible
loss alone does not identify the ratchet's mechanism. A compact results table (Table 2: setting/n,
outcome definition, headline with uncertainty, for the eight headline results) is added before the
Discussion. Reference numbering and the figure files accompany the rebuilt PDF; the bespoke unified
figures and journal-format reflow remain flagged as submission-time work.
## One point of information, not disagreement
On §2's closing remark — that conservation of an arithmetic mean's expectation is elementary and the
contribution must lie in its consequences — we agree, and would only note that the consequence now
stated (first-order cancellation of the multi-parent retention gain under output-mean inheritance,
against union-operator retention growth, in the regime where the deep tail actually lives) is the
claim we intended all along; the earlier wording claimed more than this and is gone.
Your bottom-line formulation — minimal models establish conditional results; neural experiments reveal
where the correspondences hold and break; a controlled predictive test motivates measuring functional
conflict before merging — is now, near-verbatim, how the paper describes itself. Thank you for three
rounds of genuinely improving review.

View file

@ -1,224 +0,0 @@
# Response to the external review
*Re: "The Evolution of Sex for Artificial Intelligence" (draft reviewed August 2026). This response
accompanies a revised manuscript and a set of new experiments run directly in answer to the review.
All results referenced here are committed, reproducible artifacts (configs, seeds, figures, and
per-experiment READMEs in the repository); commit-level pointers are listed at the end.*
---
Thank you for this review. It is the most useful reading the manuscript has had: it does not dispute
the programme, it disputes the *calibration* — and its central instrument, the distinction between
**interpretation, explanation, and prediction**, is exactly the right one. We have acted on it in two
ways: we revised the manuscript to claim only what the evidence supports, and we **ran the decisive
experiment the review proposed** (§5 of the review), together with four supporting experiments. The
short version of this letter: we accepted nearly everything, the manuscript is narrower and better for
it, and the prediction rung of your ladder — the one the draft "was not yet convincing on" — has now
been climbed at the small-model tier, with the falsifiers pre-registered and one internal prediction
honestly reported as unconfirmed.
## 1. The overall take, and the framing
> *"The strongest idea is … treat multigenerational model populations as systems whose inheritance,
> diversity, and compatibility must be managed—not merely as collections of models to optimise."*
Adopted, verbatim, as the stated core contribution — it now closes the abstract's first paragraph and
anchors §1. You articulated our thesis better than we had; we have taken the sentence with attribution
to the review process rather than pretend we wrote it first.
> *"The draft sometimes treats a useful biological correspondence as a mathematical identity, and an
> illustrative experiment as confirmation of a general mechanism."*
Accepted. This was the review's most consequential criticism and drove most of the textual changes
below.
## 2. Novelty: interpretation / explanation / prediction
The ladder is now explicit in §1: we state which of our claims are interpretation (merged offspring
as FisherMuller), which are explanation (the coordinate-vs-functional decomposition of merge
failure), and which are prediction. The priority-dispute language — "nobody has," "none imports,"
"the theory the tinkering has outrun" — has been removed entirely, replaced with "to our knowledge"
and positive statements of what population genetics contributes. Merge-success *prediction* is now
explicitly conceded as an occupied area (interpretable pairwise metrics; capacity/rate-distortion
accounts), with our delta stated as mechanism, not existence.
And the prediction rung is no longer only proposed — see §7 below.
## 3A. Drift, Muller's ratchet, and model collapse
Accepted in full. The manuscript now says: the *minimal inheritance model* is exactly WrightFisher;
a real learner is WrightFisher **plus a signed, architecture-specific estimator-bias operator** — and
we cite our own learning-kernel measurement against ourselves (the smoothing RNN resists collapse,
the sharpening VAE accelerates it; the drift *signs* survive in every architecture tested). Muller's
ratchet is scoped to the **irreversible arm** of collapse — the capabilities that, once lost from
every parent and source, no recombination can rebuild — with your implication stated as the reason
the correspondence earns its keep: recombination only reassembles what still survives, so the cure
must act before fixation-by-loss. The glossary entries were carrying the same identity overclaims and
have been fixed to match.
## 3B. "Merge, don't average" — the operator boundaries
Accepted. A dedicated boundary paragraph now answers your five questions in order:
- **What is conserved?** Expected rare-item mass, at the single-parent level, exact in the minimal
model's rare-item regime.
- **Under which operator?** Refitting a child to the **mean of the parents' output distributions**
that operator only. The 1/K dilution exactly cancels the K-parent union gain there.
- **Weight averaging and routing?** Explicitly labelled *empirical cousins*, not instances: a
nonlinear network's weight-mean does not compute its parents' output-mean, and a router keeps K
models' storage plus a classifier — a different parameter and inference budget from one fixed-size
child. The measured **headroom rule** is presented as the empirical bridge between the exact law
and the weight-space operators, which is all it is.
- **Does the strongest-source operator need an oracle?** Yes, and the text now says so.
- **Capacity?** When parental capabilities cannot coexist in the child's capacity, no operator
preserves the union — stated, with an explicit hand-off to the speciation section as the regime
where that boundary lives.
New supporting evidence since the review: the multi-seed replication (below) adds that fusion is not
only worse than union-preserving operators where headroom exists — it is far **less reliable**
(95% CI ±0.10 across training seeds vs ±0.026 for routing/selection), which we think sharpens the
practical half of this claim.
## 3C. Model speciation
Accepted, and this section received the most work — textual and experimental.
**Textual.** A closing block, "What these experiments do and do not establish," now states the
supported conclusion at exactly your formulation — *some merge failures reflect incompatible
functional requirements rather than a mismatch in coordinates* — and then lists the qualifiers: (i)
the conflict-condition impossibility is **information-theoretic and needs no population genetics**
(now also a formal SI proposition: endpoints and chord are invariant under any function-preserving
transformation, and any single merged model errs at rate ≥ μ(S)/2 against at least one parent); what
the genetic frame adds is locating *which divergences generate such conflicts*; (ii) the
epistasis-positions-the-cliff claim and the snowball are labelled **hypotheses at the neural tier**,
verified only in the analytic model; (iii) alignment claims are scoped to the enumerated symmetries
of the architecture tested, and "unmergeable" means by aligned linear interpolation — a barrier to
that operator does not preclude every recombination method (routing sidesteps it by not blending).
Emergent DobzhanskyMuller incompatibilities are carried as the flagship *hypothesis*, with the
regimes where our tests found none stated as bounds.
**Experimental (new since the review).**
1. *Alignment under the full symmetry group.* Anticipating the "one control does not prove the
optimum over all allowed symmetries" objection — and the 2026 richer-symmetry results — we
re-ran the decomposition aligning modulo the **complete** function-preserving unit symmetry group
of the ReLU MLP (per-unit positive rescaling ∘ permutation; the aligner provably recovers a
permuted-and-rescaled copy exactly). The conflict residual is unchanged (0.502 → 0.497); the
independent-init barrier still vanishes (0.001). The cliff now also carries a hybrid-fitness
readout: merged accuracy 0.97 → 0.03 with conflict.
2. *The emergent test, pre-registered.* Divergent-but-compatible specialists (disjoint classes;
shifted-view conventions), out to 6.4× the base training: residual 0.000 everywhere, and the merge
*rescues* the two forgetting parents (~0.50 → 0.955). We report this null prominently — you
identified the sharper question ("which kinds of specialisation create merge-breaking
interactions, and which remain complementary?") and this is its first half of an answer:
*specialisation on shared ancestry did not break merging in any regime we tested; imposed
functional conflict always did.*
3. *The LLM tier.* The same two knobs in 0.5B LoRA children of a frozen base (which pins the
coordinate system, so merge failure is functional by construction): conflicting conventions
produce **function-specific** hybrid breakdown (merged coherence below both parents; private,
disjoint skills unharmed in a budget-controlled design, 3 seeds), and over-training disjoint
specialists 1→12 epochs produces **no** emergent isolation (the merge improves, staying above the
best parent).
## 4. Importance, scope, and the supporting overstatements
- **"Three task families and one seed."** The LLM claims are now multi-seed with fixed test sets:
merges beat every specialist with non-overlapping CIs on the sharper metric (5 seeds); union vs
fusion and directed-selection vs soup replicated at 3 seeds on the hard benchmark. Three
lexically-distinct families remain a stated limitation; the full grounded LLM society remains
explicitly unbuilt and is flagged as such.
- **Open-ended growth.** Accepted — the §11 closing has been rewritten: the architecture removes the
*storage* obstacle to indefinite accumulation; that is bookkeeping, not a demonstration of
unbounded capability growth, which our deliberately finite models do not test.
- **Frozen base ≠ unchanged behaviour** — fixed (§3 now guarantees a recoverable core, not
unchanging conduct).
- **Baldwin effect** — now an *echo*, with the mechanism difference stated (selection for genetic
assimilation vs direct distillation).
- **Consolidation and the archive** — fixed: a digital system can and should keep every ancestor;
the irreversibility is *operational* (nothing in the production loop consults the archive by
default), and the safeguard now includes an audit that diffs against the archived ancestor.
- **"Control theory" → "framework"** throughout, subtitle included.
- **The claimassumptionsevidencelimitation table** is in §13 ("The claims at a glance"), ten rows,
each labelled exact / empirical / hypothesis with known limits.
## 5. The decisive experiment — run
We implemented your six-step design as specified, at the 0.5B tier (39 parent pairs, 3 seeds, fixed
held-out test sets, falsifiers pre-registered in the config before running):
1. **Controlled interaction structure:** three axes decorrelated by construction — *conflict*
(contradictory conventions on shared ambiguous prompts, private budgets fixed), *compat* (the
same shared prompts learned under the **same** convention: overlap without conflict), and
*duration* (weight divergence with zero conflict, 1→12 epochs).
2. **Functional divergence separated from duration and weight distance:** the duration and compat
axes span the same weight-divergence and data-volume ranges as the conflict axis, at ~zero merge
penalty.
3. **Operational epistasis, pre-merge:** confidence-weighted bilateral disagreement on a probe mix
drawn blind to where the conflict lives — the theory's point being that raw disagreement counts
harmless *complementation* (one parent ignorant) as conflict, while the DobzhanskyMuller
structure is *bilateral confident contradiction*.
4. **Against existing predictors:** gradient alignment at the shared base, LoRA-delta cosine and L2
(computed exactly), and a performance-based (cross-family accuracy) baseline.
5. **Operator choice under matched budgets:** partially — see honest riders.
6. **Held-out tasks, multiple seeds:** yes (fixed tests, 3 training seeds).
**Result.** Against the pre-registered primary outcome (merge penalty = parent potential merged
achieved, the hybrid-load analogue): functional measures predict (raw disagreement ρ = +0.46,
operational epistasis ρ = +0.45, both p < 0.005); gradient alignment is weakly informative (0.35);
**both geometry predictors are uninformative** (delta-cosine +0.03, delta-L2 +0.17, n.s.);
performance-based ~0. *Functional conflict, measured before merging, predicts merge failure; weight
divergence does not.*
Two things about how this result was reached that we want on the record:
- **The control that broke our own experiment first.** In the initial two-axis grid, the *best*
predictor was delta-cosine (ρ = +0.60) — geometry appeared to win. We identified the confound
(every shared-data pair in that pool was a conflicted pair, so geometry could succeed as a mere
overlap/volume detector), added the compat control axis, and geometry's correlation collapsed to
+0.03 while the functional measures held. We report this sequence in the results README rather
than presenting only the final table.
- **An internal prediction failed, and we say so.** We pre-registered that confidence-weighting
should beat raw disagreement as a rank predictor. It does not (they are statistically
indistinguishable at n = 39); the weighting does double the conflict-vs-compat contrast in levels
(2.0× vs 1.5×). The paper reports the functional-vs-geometric verdict, not a win for the
refinement.
**Honest riders:** correlations are moderate (|ρ| ≈ 0.45), bounded by the large intrinsic seed
variance of 0.5B weight-averaging (itself now a documented finding); your step 5 (operator choice
under matched budgets) is only partially delivered — the soup-vs-route gap readout is
noise-dominated at this scale; and the whole result is one model family at one scale. The 7B
replication on the HPC cluster is the planned firm-up before we treat this as more than a
small-model demonstration.
## 6. Where we (mildly) push back
Only two places, both narrow. First, on *"the impossibility does not require population genetics"*
agreed, and now stated; but we would defend the framework's role in the surrounding structure: it
told us *which* pre-merge measurement to make (bilateral confident contradiction rather than raw
disagreement or distance), *which* control to build (complementation ≠ conflict), and *which* null to
pre-register (emergent isolation) — and those choices are what the decisive experiment's outcome
vindicated against the geometry baselines. Second, on *"union preservation risks being built into the
operator's definition"* — the conservation law's content is the exact *failure* of the mean operator
(the 1/K cancellation), not the definitional success of the max operator; we have tried to make the
text carry it that way, with the oracle requirement explicit.
## 7. What we have not done
The full grounded, diversity-preserving multigenerational LLM society (still the stated largest gap);
7B replication of the decisive experiment; an entanglement measure for *real* task pairs (our
epistasis knob is constructed); the operator-choice decision test at usable signal-to-noise; and
ambiguous/overlapping task families where routing stops being trivially easy. These are listed in the
manuscript's open-problems section in this form.
## Changelog
Manuscript revision: commit `58e6c74` (claim-narrowing; all §1§4 and draft-level items above).
New experiments: `ea051a5` (full-symmetry alignment + emergent null, MLP tier), `5a23dda` (LLM-tier
speciation + multi-seed replication), `287d232` (the decisive experiment + its control axis). The
revised manuscript is `paper/the-evolution-of-sex-for-ai.md`; per-experiment analyses are in
`results/*/README.md`; every figure regenerates from committed artifacts.
We would welcome another pass — in particular on whether the decisive experiment's design and its
riders are stated at the right strength, and on whether the remaining hypothesis labels
(epistasis-cliff and snowball at the neural tier; emergent DMIs) are placed where you would place
them.

View file

@ -1,304 +0,0 @@
# The Lamarckian Society — Summary of results
*The complete laptop-reproducible body of work: the analytical core (Layer 1), the
architecture-general neural existence proof (Layer 1.5, including real MNIST), the learning
kernel, and the sexual-reproduction society (E7E11). Two summaries of the same work — one
technical, one accessible.*
**The arc in one breath.** Model collapse is **asexual, self-consuming degradation**: a lineage
trained on its own outputs drifts to its own mode and loses the rare tail. We show this is
*literally* a WrightFisher drift process (validated against closed forms), reproduce it in real
trained neural weights and on real MNIST images, and then establish the **cure** — a **grounded
sexual society**: reality-checking (grounding) plus **recombination across many decorrelated
parents** (sexual reproduction, not teacher→pupil copying) plus **quality-diversity** selection.
The payoff is not merely arrested collapse but a population whose **offspring exceed their
parents** and whose general capability **climbs** while specialties are re-earned — and removing
any one operator breaks it, and a first **real-LLM prototype** (up to 7B on HPC) confirms the
recombination claims in trained weights. **131 tests pass**; three of the core predictions are exact
closed forms, so the headline curves sit on analytic targets rather than eyeballing.
---
## A. Technical summary
### 1. The analytical core — collapse as WrightFisher drift (Layer 1)
Knowledge is a distribution `p_t` over `K` items on the simplex; a fixed Zipf-tailed truth `p*`;
the generational step — *sample `n` from the parent, mix in `m` fresh real samples, refit* — is
**literally a WrightFisher process with immigration**, not an analogy. Each safeguard from the
perspective paper is one operator on that step (grounding `g=m/(n+m)`; region-matched grounding;
multi-teacher recombination; directional vs quality-diversity selection; re-minting). Because the
process is WrightFisher it inherits **closed-form validation targets**, enforced as
`<0.5%`-tolerance assertions (the "spine of trust"): neutral heterozygosity decay
`E[H_t]=H₀(11/n)^t`; fixation probability = initial frequency; the *exact* mutationdrift
equilibrium `H_eq = H*·m(2n+m1)/(n+2nm+m²)`; the tail-persistence threshold `m·p*ᵢ ≳ 1`; and the
recombination union coverage `U(K_T,ρ,q)=T[ρq+(1ρ)(1(1q)^{K_T})]`.
Findings **E1E6**:
- **E1 — collapse (null).** Neutral drift reproduces the geometric `H` decay to Monte-Carlo error;
support collapses `K→1`; forward-KL diverges; tail *items* die ≈10× faster than head items.
(Aggregate tail *mass* is a drift martingale — a misleading metric; tail-*item* survival is the
honest one.)
- **E2 — grounding phase boundary (headline).** Stationary `H` tracks the exact `H_eq`; a critical
**`g* = 0.048` (CI [0.047, 0.050]) ≪ 1** — as little as one real sample against 200 inherited
restores ~68% of the truth's diversity; `g=0.05` reaches 96%. The sharp threshold lives in
discrete tail-item survival, not the smooth `H`. The **deep tail is unrescuable** by grounding at
feasible budgets (`m* 1/p_min`) — which is what recombination is for.
- **E3 — region-matched grounding.** At fixed budget, matched grounding holds the exercised region's
tail (0.49) where uniform lets it collapse (0.07). Grounding protects only what it overlaps.
- **E4 — multi-teacher recombination ("merge, don't average").** Union coverage matches the closed
form exactly (recombination *supplies* the tail). **Principal finding:** under mean-mixture
distillation surviving tail coverage is **flat in `K_T`** — a conservation law (averaging's `1/K_T`
dilution exactly cancels the union gain). The benefit is realised only under a **union-preserving
merge** (`max` over teachers, à la M2N2 model-merging). *Merge weights; don't average outputs.*
- **E5 — QD vs greedy.** Greedy selection fixes (`H≈0.01`); quality-diversity (`w_i ∝ f_i·p_i^{α}`)
holds `H` at a positive plateau (0.480.88, rising with novelty `α`).
- **E6 — re-minting gate.** Re-minting a *collapsed* lineage makes forward-KL to the original
diverge (irreversible lock-in); a diversity gate (`H≥H_gate`) prevents it; healthy re-mint is
harmless.
### 2. Collapse in real trained weights, and on real images (Layer 1.5)
A re-scoped, cheaper Layer 2: realise the *same* WrightFisher abstractions in **real trained
generative models** on a fully-synthetic sandbox with an exact oracle, then confirm on real MNIST.
A model's "knowledge" is its oracle-measured distribution over `K` modes; the generational step is
*train-a-model-on-the-previous-model's-samples + grounding*.
- **The histogram bridge (HARD GATE).** A memoryless histogram model reduces Layer 1.5 *exactly* to
Layer 1: run through the neural runner it recovers `g* = 0.047` and sits on the exact `H_eq` curve.
This licenses every trained-model result to be read against the analytic core.
- **Collapse in an RNN, and the metric reframing.** A GRU retrained each generation on its own
output drifts from truth (forward-KL climbs) and grounding arrests it — the sign confirmed. But
the **operative neural collapse metric is forward-KL, not `H` or tail-survival**: the RNN's
*smoothing* inductive bias keeps spurious tail modes alive, so `H` stays ~80% of `H*` and
tail-survival is non-monotone in `g`. On forward-KL, half the divergence gap closes by a
median-recovery grounding `g≈0.04` (echoing Layer-1's 0.048), but *full* recovery needs `g≈0.19`
the sharp `g*≪1` is an exact-operator feature the trained net *softens*.
- **Architecture-generality.** Collapse + grounding-rescue appear in the **histogram, GRU, and MLP**
alike — the operator is not an artefact of one model class.
- **Recombination in real weights.** The E4 "merge, don't average" law reproduces: construction-level
union rises 0.49→0.96, oracle-guided max-merge surviving coverage rises while mean-distill stays
flat — the conservation law holds in trained weights (compressed/noisier, the expected smoothing
caveat).
- **Real-MNIST external validity.** A convolutional VAE (the canonical collapse vehicle) retrained on
its own generated digits, modes = digit-class × stroke-thickness (K=30, Zipf), read by a frozen
CNN oracle (98.5% mode accuracy, confusion matrix recorded as the noise floor): the **dry lineage
collapses to a single mode** (forward-KL 0.5→18, support 30→1, tail wiped out, H→0), while **10%
grounding holds all 30 modes**. The eyeball montage shows varied gen-0 digits degenerating into one
blurry blob. Collapse and its cure are real on real images — not a synthetic artefact.
### 3. The learning kernel — neutral drift is a null both neural models fail, oppositely
Prompted by revisiting the neural deviations, the refit step is generalised from a pure resample to
`p_{t+1} = T_θ(counts/n)`: a **learning kernel** with a *smoothing* knob (mutation toward a prior)
and a *sharpening* knob (mode-competition), both identity by default (so Layer 1 is untouched).
Result: **neutral WrightFisher fails both neural architectures in opposite directions.** The
**VAE** (large `n`, small `K`): neutral drift is *inert* (no collapse), yet the real VAE collapses
to one mode — a **sharpening** kernel reproduces it (the estimator *adds* collapse pressure). The
**RNN**: neutral drift drives `H→0`, but the real RNN only partially collapses — a **smoothing**
kernel reproduces the `H`-floor (the estimator *removes* collapse pressure). Model collapse in real
learners = **neutral drift ⊕ an architecture-specific, signed estimator-bias operator**; this
mechanistically explains the architecture-generality result and the softened neural `g*`.
### 4. The sexual-reproduction society (E7E11) — from teacher→pupil to sex with unbounded parents
The single-locus, fixed-`p*` world can only express *recovery toward a ceiling*. The society's
load-bearing claim is **vertical** — capability that *exceeds* any component — which needs
combinatorial structure. Knowledge becomes a distribution over **genotypes** (`L` biallelic loci,
fitness = number of correct loci), and the one new operator is **recombination**. This is where the
frame shifts: **teacher→pupil distillation is asexual copying (caps at the ceiling); recombination
is sexual reproduction (combinatorial, generative — offspring can exceed both parents), and unlike
biology there is no two-parent limit.** The celebrated evolution-of-sex theory maps onto the thesis
exactly (collapse = Muller's ratchet; merging = meiotic reassortment; "exceeding" = the
FisherMuller effect):
- **E7 — the advantage of sex.** A population adapting toward an optimum: the **sexual lineage adapts
faster** than the asexual one (clonal interference), keeping loci in linkage equilibrium (LD→0 vs
an LD spike). Honest scope: a *speed* advantage, not a permanent gap (the single-population ratchet
is subtle).
- **E8 — the vertical claim (FisherMuller), the headline.** Decorrelated *parents* are specialists
(expert on their loci, agnostic elsewhere). **Sexual recombination assembles a genotype fitter than
any parent, climbing to the optimum (12/12 — a genotype no parent had)** as parent count grows and
correlation `ρ→0`, while the best single parent (~8.7) and the mean-mixture "model soup" (~11.6)
plateau below.
- **E9 — landscape robustness ("why sex?").** On rugged (epistatic, NK) landscapes, blindly
recombining trained models causes **outbreeding depression** — offspring fall *below* the parents,
worse the more entangled the skills and the higher the recombination rate — and the **optimal
recombination rate shrinks as ruggedness grows**. Design rule: *merge freely when skills are
complementary; sparingly, and with selection, when entangled.*
- **E10 — directed sex beats biological sex (the AI superpower).** Biology is stuck with two
random-mating parents and no offspring preview; an AI can **choose complementary mates, evaluate
many recombinant offspring, keep the fittest, over rounds, with unbounded parents**. Random
("biological") sex craters with ruggedness (0.66→0.51, deep outbreeding depression); **directed sex
tracks or exceeds the best parent at every ruggedness** — a catastrophe turned into a win, with no
biological analog.
- **E11 — the dynamic Lamarckian society (the C3 vertical claim, realised).** A finite population of
agents (genotypes) evolves on a rugged NK landscape that *is* reality, composing all four operators
— grounding, directed sex, quality-diversity, mutation. Grounding is made load-bearing by the
**consensus-conformity (self-consumption)** mechanism: selection acts on
`g·true_fitness + (1g)·conformity`, so at `g=0` the society optimises fitting-the-crowd rather than
reality. A **4-arm ablation (global optimum ≈ 0.79), each breaking distinctly, only the full society
climbing:** `full` **0.78** (climbs to the optimum, diversity maintained longest) · `no_sex` 0.77
(can't recombine to escape local optima) · `no_diversity`/greedy 0.74 (collapses diversity fastest,
stuck at a worse local optimum) · **`no_grounding` 0.48** (self-consumption collapse to an unfit
consensus — trains on the crowd, regresses to a confident-but-wrong mean; the population strongly
*agrees* while being *wrong*). The society needs **all** of grounding + directed sex + diversity: on
a rugged landscape you need diversity to explore basins, sex to recombine them, and grounding to
select on reality.
### 5. The claims tested in real LLM weights (Layer 2 prototype)
The analytic operators above make three claims that are testable in *real language-model weights*:
"merge, don't average" (E4), the FisherMuller generalist-from-specialists (E8), and directed sex
(E10). We tested them with a small, reproducible pipeline — LoRA **specialists** fine-tuned on three
*disjoint*, procedurally-generated task families (`lists`, `strings`, `arith`) with an **exact-match
verifier** as reality's "no" — recombined by various operators and scored on held-out tasks. Runs are
at two scales: **Qwen2.5-0.5B** locally and **Qwen2.5-7B** on Imperial College's CX3 HPC (one L40S).
This is a prototype (three families, one seed), so read it as **signs, not magnitudes**; the analytic
layer carries the quantities.
- **Recombining specialists yields a generalist that exceeds every parent (E8).** At 7B, a weight-space
**merge of the three specialists reaches 0.87 overall vs 0.77 for the best single specialist**, and
beats every specialist on every family — the FisherMuller signature, in real weights. At 0.5B the
same sign is present but marginal, because a weak base *dilutes* (see below).
- **Union beats averaging — when there is headroom (E4).** Two ways to recombine: **fusion** (average
the LoRA deltas — a "model soup") versus **union** (keep each specialist intact and *route* each
input to the right one — a Mixture-of-Experts). On tasks with headroom, **routing beats fusion**
because averaging dilutes a specialist's contribution: at 0.5B routing 0.74 > soup 0.64; on *hard*
7B tasks routing 0.50 > soup 0.40, with fusion diluting the fragile string-cipher specialist so badly
(0.67→0.30) that the soup even loses to the best single specialist. This is E8's `max > mean` in real
weights.
- **Directed sex helps — when there is headroom (E10).** Breeding a population of recombinant offspring
(specialists merged at many weightings), scoring each on a held-out split with the verifier, and
keeping the fittest **beats the single a-priori soup** where the soup is suboptimal: 0.69 > 0.64 at
0.5B, and 0.49 > 0.39 on hard 7B tasks — recovering most of routing's benefit from one deployable
model. This is offspring selection (which biology cannot preview) realised in weight space.
- **The one law that unifies them: headroom, not scale.** Our first 7B runs on *easy* tasks showed the
opposite — fusion beating union, and directed selection tied with the soup. That was a **saturation
artefact**: the easy families saturated 7B at 1.00, leaving no room to lose to dilution, so the naive
soup was already optimal. On *hard* (unsaturated) 7B tasks the original ordering returns. **The
operative variable is headroom**: union-preserving merging and offspring selection pay off whenever
there is room to lose to dilution — a weak base *or* hard tasks — and only the degenerate corner of
easy tasks on a strong base makes naive averaging look sufficient. This both confirms E4/E8/E10 in
real weights and delimits exactly when the refinements matter.
Honest scope: three lexically-distinct families (so the router is trivially accurate — an ambiguous-
skill benchmark is the interesting next stress test), one seed, small LoRA. The full *grounded*
society on LLMs (collapse under dry self-training, the dynamic society) is the HPC-scale next step.
### 6. Positioning — what is prior art, what is ours
An independent 2026 paper (Riis, *Drift and selection in LLM text ecosystems*) rigorously formalises
**collapse = WrightFisher drift** (martingale of minority mass, rare-first extinction, de Bruijn
fixed points, drift+selection) with n-gram agents. **We concede that framing as prior art and cite
it — "collapse is drift" is no longer our contribution.** Crucially, Riis's "mixed environment"
retains the lineage's own *old synthetic* tokens (no injection of external truth), so his headline is
*pessimistic* (extinction is independent of retention). **Our defensible contributions, ranked:**
(1) **recombination as sexual reproduction** — the "merge, don't average" law, the FisherMuller
vertical claim, directed sex, and their limits (E4, E8E10) — an operator Riis lacks; (2) the
**learning-kernel / estimator-bias axis**, which he *names as future work*; (3) **grounding as
immigration from a fixed reality**, giving a critical `g*≪1` his closed loop cannot have; (4)
**architecture-generality in real weights + real images (MNIST)**; and (5) **the integrated dynamic
society and its vertical claim** (E11), wholly ours. The repositioning: from *"collapse is drift"*
(diagnosis) to **a population-genetic control theory for sustaining open-ended knowledge** — the
engineered cure and its integration.
### Design rules that fall out
1. **Never inherit dry, and ground where it matters** — a little reality (`g*≈5%`) protects most
diversity, but it protects only what it overlaps, and it can't hold the deep tail.
2. **Merge, don't average — where there is headroom** — union-preserving model-merging (routing /
max-merge) realises the multi-teacher benefit; averaging dilutes it. Confirmed in real LLM weights,
with a sharp caveat: dilution only bites when the task leaves room for it — on easy tasks a strong
base's naive soup already composes to the ceiling, so averaging looks fine. The refinement matters
exactly in proportion to how far the naive soup is from optimal.
3. **Sex, with no parent limit** — recombining decorrelated specialists yields offspring that exceed
any parent; more, complementary parents climb higher.
4. **Match recombination to entanglement** — merge freely for complementary skills; sparingly for
entangled ones; and *always select offspring* (directed sex), which AI can afford and biology
cannot.
5. **Keep diversity, and stay grounded** — greedy selection or a broken reality-signal both collapse
the society; only grounding + sex + diversity together climb.
6. **Gate irreversible consolidation on diversity** — re-mint a base only while the lineage is
healthy.
---
## B. Accessible summary (for ML engineers and neuroscientists)
### The question
Modern AI is trained once and frozen because it cannot keep learning without *catastrophically
forgetting*. The Lamarckian Society proposes an alternative: **generations** of bounded agents that
learn, then reproduce — passing on what they learned. But there is a notorious trap: train a model on
the previous model's outputs, generation after generation, and it suffers **model collapse** — the
rare cases (the *tail*) vanish first and the model drifts to a bland mode. So the whole scheme lives
or dies on one question: **when does generational transmission accumulate knowledge instead of
degrading it?**
### The through-line: collapse is *asexual* degradation; the cure is *sex*
The key reframe of this work is that **teacher→pupil copying is asexual reproduction** — a pupil can,
at best, recover what its teachers had. That caps out, and left alone it degrades (collapse). The cure
is **sexual reproduction between agents**: combine *many decorrelated parents* so the offspring
inherits a *combination* none of them had — and can be **better than any parent**. Unlike biology,
AI sex has **no two-parent limit** and can *choose* mates and *select* offspring. That is the engine
that lets a society climb instead of collapse.
### What we found, in plain terms
1. **Collapse is real, and it's math.** Casting generational training as the century-old
**WrightFisher** drift process (not a metaphor — the same equations) gives exact formulas to
check against. The rare stuff dies ~10× faster than the common stuff.
2. **A little reality goes a long way — but not for everything.** Mixing in even ~5% verified real
data restores most of the diversity and holds it (**grounding**). But the *very rarest*
capabilities can't be saved by grounding alone — that's what recombination is for.
3. **Collapse is real in actual neural nets, and on real images.** We reproduced it in trained RNNs,
MLPs, and a VAE, and on **real MNIST** — where a VAE trained on its own digits collapses to a
single blurry blob, while a little grounding keeps all the digit styles alive. (Honest nuance: real
nets *smooth*, so "how many modes are alive" lies to you; "how far from the truth" is the honest
ruler.)
4. **To fight tail collapse with many teachers: merge, don't average.** Averaging their outputs
mathematically cancels the benefit; a *merge* that keeps each item's strongest source realises it.
5. **Sex makes offspring that beat their parents.** Recombining decorrelated specialist models
assembles capabilities none of them had, climbing to the optimum as you add more, complementary
parents — while averaging ("model soup") and the best single parent plateau below. This is a
celebrated evolutionary result (the FisherMuller effect), now shown for AI model merging.
6. **But sex can backfire — and AI has a fix biology lacks.** When skills are *entangled*, blindly
merging good models produces *worse* offspring ("outbreeding depression"). The fix is **directed
sex**: choose complementary partners, generate many merges, and keep the best — which AI can do
and biology can't. Directed sex turns the catastrophe into a win.
7. **The whole society climbs only with all the pieces.** In an evolving population on a rugged
"reality" landscape, the *full* society (grounding + directed sex + diversity) climbs to the top
while keeping its specialists; remove **grounding** and it collapses into a confident, wrong
consensus (the exact analogue of training on the internet's AI-generated crowd); remove **sex** and
it gets stuck; remove **diversity** and it converges too fast to a worse answer. Each failure is
distinct; only the full society wins.
8. **It shows up in real language models — with one clean caveat.** We merged LoRA-specialised
Qwen models (0.5B locally, 7B on a university GPU cluster): the recombined model beats every
specialist (FisherMuller, for real), and *routing* / *offspring-selection* beat naive averaging —
but only when the tasks are hard enough to leave room. On easy tasks a strong model's plain average
is already at ceiling, so the fancier operators don't help. The lesson is precise: **these
recombination tricks matter exactly in proportion to how far the naive average is from the best you
could do** — which is a genuinely useful thing to know before you spend compute on them.
### Why it is novel and why it matters
- **It turns model collapse from a warning into a control theory.** Collapse-as-drift is now known
(and independently formalised elsewhere). Our contribution is the *cure* and its integration: a
grounded, sexually-reproducing, diversity-preserving society that not only avoids collapse but
**climbs, with offspring exceeding parents**.
- **The sexual-reproduction frame is, we believe, genuinely new for AI** — model merging reframed as
meiotic recombination, with a rigorous account of when it helps (complementary skills), when it
hurts (entangled skills), and how to make it reliably win (directed sex, unbounded parents).
- **It is validated, not vibes.** Three core predictions are exact closed forms; the neural, image,
and **real-LLM** results confirm the *signs* in real trained weights (up to 7B on HPC); 131
automated tests pass; the whole study is laptop-reproducible from a seed (the LLM tier statistically
reproducible on one GPU).
- **It gives concrete design rules** for anyone building self-improving or model-merging systems:
ground where it matters, merge-don't-average, match recombination to skill-entanglement, select
offspring, keep diversity, and gate irreversible consolidation on health.
*The remaining frontier is the **LLM instantiation** — realising the grounded sexual society with
actual language models (LoRA specialists, real model merging, execution-verified grounding), which
the blueprint frames as the eventual empirical rung.*

View file

@ -1,74 +0,0 @@
# SI notes — drafts of formal statements for the PNAS manuscript
*Working drafts; folded into the SI Appendix at Phase 4. Each statement is written to be exactly as
strong as what is true — no more.*
## S1. The incompatibility floor: what no alignment can remove (E13c)
**Setting.** Models A and B are trained on the same input distribution; their target label functions
`f_A` and `f_B` agree except on a conflict set `S` of probability mass `μ(S)` (in E13's conflict
condition, the cyclically-relabelled classes; `μ(S) ≈ conflict_frac` up to class balance). A
*function-preserving transformation* `T` (any composition of hidden-unit permutations and, for ReLU
networks, positive per-unit rescalings — the full unit symmetry group of a plain ReLU MLP) satisfies
`T(B)(x) = B(x)` for all `x` by construction.
**Proposition 1 (endpoint invariance — with the term "chord" defined precisely).** Here "chord"
means the α-linear interpolation **of the endpoint loss values**, `(1α)·L(A) + α·L(B)` — the
baseline in the barrier definition, a function of the endpoints only — NOT the weight-space
interpolation path. For every function-preserving `T`, the endpoint functions, hence the endpoint
losses and this chord, are identical for `(A, T(B))` and `(A, B)`. The **interpolation path itself is
generally NOT invariant** — losses along `(1α)·A + α·T(B)` change with `T`, which is precisely why
alignment can lower a barrier. *(Immediate from the definition of function-preserving.)* Scope
caveat: our aligner provably recovers a permuted-and-rescaled copy exactly — an important special
case — but this does not establish global optimality of the alignment over the symmetry group for
independently trained networks; the decomposition's "removable" share is therefore a lower bound, and
the "residual" an upper bound, on their true values.
**Proposition 2 (no merged model can serve both parents).** Let `h` be *any* single classifier (in
particular, any interpolated/merged model, under any alignment). On every `x ∈ S`, `f_A(x) ≠ f_B(x)`,
so `h(x)` disagrees with at least one of them. Hence
`ε_A(h) + ε_B(h) ≥ μ(S)`, and therefore `max(ε_A(h), ε_B(h)) ≥ μ(S)/2`,
where `ε_P(h)` is `h`'s error against parent `P`'s labels. A hybrid of two models whose conventions
conflict on mass `μ(S)` errs at rate at least `μ(S)/2` against at least one parent — **hybrid
disadvantage with an information-theoretic floor, independent of the alignment group, the
architecture, and the merging operator.** This is reproductive isolation in the fitness sense: past a
given functional conflict, *no* recombination operator produces an offspring loyal to both lineages.
**What remains empirical, and why the experiment is designed as it is.** Propositions 12 do *not*
bound the single-task path barrier (the loss along the interpolation between A and `T(B)` evaluated
on one parent's task): in principle a path could dip toward one parent's function. Whether it does is
exactly what E13 measures — and the measured answer is that it does not: the conflict-condition
barrier is unchanged by permutation alignment (`residual`) *and* by alignment modulo the full
permutation × positive-rescaling group (`residual_scale`), while the same aligner removes ~all of the
independent-init barrier (the positive control). Richer-symmetry results for transformers
(arXiv:2606.23607; neuron-identifiability approaches to linear mode connectivity, 2026) strengthen
the *removable* side of the decomposition and are therefore complementary: the more barrier a larger
group can remove for *compatible* models, the sharper the meaning of the residual that survives for
*incompatible* ones — and Proposition 2 caps what any of them could ever achieve on the conflict set.
**Terminology note for the paper.** "Residual (after alignment)" = the estimated functional
incompatibility; for ReLU MLPs we align modulo the full unit symmetry group, so the estimate is not
confounded by missed symmetries of that architecture class.
## S2. Emergent vs imposed incompatibility (E13b framing)
The conflict condition *imposes* contradiction (the two label maps disagree on `S`), which pins
`μ(S) > 0` and activates Proposition 2. A true BatesonDobzhanskyMuller incompatibility is
*emergent*: each lineage's substitutions are harmless on their own background (`μ(S) = 0` — the
training signals never contradict), and incompatibility, if any, arises only in the *combination*.
The `disjoint` (complementary class specialists) and `augment` (divergent input conventions)
conditions realise this: any residual barrier they develop cannot be attributed to label conflict and
is the emergent-speciation signal proper. Pre-registered readings: residual grows with divergence →
model speciation is emergent in real weights (E12's trajectory realised); residual stays at the
`shared`-control level → within this regime, trained networks are *more* merge-compatible than the
biological analogy predicts — an honest bound on the analogy, and itself a design-relevant result
(merging is safe absent functional conflict).
**Outcome (2026-08-11 run, 4 reps, t_div ≤ 3200): the second reading.** Residual 0.000 at every
divergence in both emergent conditions, and the merge *rescues* the forgetting `disjoint` specialists
(parents → 0.535/0.474 on the full task; merged ≈ 0.955 throughout — a sustained FisherMuller rescue
at zero barrier). Isolation in real weights required functional conflict in this regime; whether
long-horizon over-specialisation erodes mergeability at LLM scale (cf. arXiv:2607.11997) is the
`llm_speciation` question (Phase 3).

View file

@ -1,529 +0,0 @@
# The Evolution of Sex for Artificial Intelligence — the plain-language version
### How ideas from breeding and genetics tell us how to keep AI models improving across generations
*This is an accessible companion to the full paper (`the-evolution-of-sex-for-ai.md`). It makes the
same argument and reaches the same conclusions, but assumes only that you know roughly what a machine
learning model is — that it is trained on data, that training adjusts numbers called "weights," and
that you can fine-tune a model on new data. Everything else is explained as we go. Where the full
paper defends each point against the research literature, this version just tells the story.*
**Giorgio F. Gilestro** · Department of Life Sciences, Imperial College London ·
giorgio@gilest.ro · https://lab.gilest.ro
---
## The one-paragraph version
If you train an AI model on the output of earlier AI models, over and over, it rots: rare knowledge
disappears and everything drifts toward a bland average. This is a known problem ("model collapse"),
and in its simplest form it is governed by *exactly* the same math that describes how small biological populations lose
rare genes by chance. That is bad news, but it comes with good news: biology has been managing this
kind of rot for hundreds of millions of years, and its best-tested remedy is **sex** — making offspring
by *combining* several parents instead of copying one. This paper tests how far that remedy carries
for AI. This paper takes ninety years of genetics about
*when and why sex beats cloning* and reads it as an engineering manual for building AI that keeps
getting better across generations instead of decaying. Along the way it produces concrete, testable
rules — including a surprising one about *how* to combine models ("merge, don't average"), and a limit
("models can drift so far apart they can no longer be usefully combined at all"). We back the argument
with small, fully reproducible experiments and a first test on real language models.
---
## A few words you'll need
- **Model collapse** — what happens when you train models on the output of earlier models, again and
again: rare cases vanish, everything gets blander. The central disease this paper is about.
- **Fine-tuning / specialising** — taking a trained model and training it a bit more so it gets good at
one specific thing.
- **Model merging** — combining two or more trained models directly, by mixing their weights, to get
one model — *without* retraining. Think "breeding two models" rather than "teaching a third."
- **The tail** — the rare stuff. Common knowledge is the "head" of the distribution; unusual cases,
rare facts, and edge behaviours are the "tail." Collapse eats the tail first.
- **Grounding** — mixing some real, verified data from the actual world into training, instead of only
model-generated data. The reality check.
- **Genetic drift** (from biology) — in any finite population, rare gene variants can vanish purely by
chance, because not every individual reproduces. This is the biological twin of model collapse.
- **Recombination / sex** (from biology) — making a child by combining pieces of more than one parent.
The opposite of cloning (**asexual** reproduction).
---
## 1. A society of AIs across *time*, not just space
When people imagine "many AIs working together," they usually picture teamwork in the *moment*:
several specialist agents side by side, splitting up a job. This paper is about a different direction:
**time**. Not AIs that cooperate right now, but AIs that pass knowledge down across **generations**
each new model starting from what the previous ones learned, the way each human generation inherits
the accumulated knowledge of the last and adds a little.
The key event, then, is **reproduction**: making a new model out of older ones. A single model, like a
single person, eventually stops improving. A *lineage* — a chain of models across generations — does
not have to. Civilisation isn't smart because any one person is; it's smart because knowledge
accumulates. The whole question of this paper is: **how do you make one AI model out of older ones,
without the knowledge rotting on the way down?** That's exactly where it can go wrong.
(This corner of AI is suddenly busy: several 20252026 research projects build populations of models
that improve over rounds, and "model merging" has become a small industry that already borrows words
like crossover, mutation, and mate choice. What this paper adds is the quantitative framework behind
those borrowed words — and honest tests of where it works and where it doesn't.)
## 2. Why today's models can't do this
Today's large models have no life cycle. A model is trained once, at huge expense, then **frozen** and
shipped. It does not learn from the people who use it. "Learning" and "doing" are two separate eras
with nothing connecting them.
There's a real reason for the freeze: if you keep training a neural network on new things, it tends to
overwrite what it already knew. This is called **catastrophic forgetting**, and it's been a known
problem since the 1980s. Freezing dodges it by refusing to learn at all. But a lineage needs the
opposite of a frozen model: it needs members that keep learning through their working lives and then
pass on what they gained. So step one is a model that can *grow safely.*
## 3. A model that grows without forgetting
The trick is to stop overwriting. Keep the model's original core frozen and untouchable, and bolt each
new skill onto *extra* capacity added beside it. In practice this is what small add-on "patches" like
**LoRA** already do: the big pretrained model stays fixed, and you train a little attachable module for
each new skill. If the core is never altered, its knowledge can't be erased — though the system's
*behaviour* can still change while patches are active; what's guaranteed is a recoverable core.
There's even a rough brain analogy: we have a fast memory (the hippocampus) that grabs an experience
immediately, and a slow memory (the cortex) that absorbs patterns gradually, usually while we sleep.
The AI version is clean: the prompt is short-term memory, a database is fast memory, the trained
weights are slow memory, and a periodic "consolidation" step moves knowledge from fast to slow.
One consequence matters a lot: because this kind of model only ever *adds* capacity, it eventually
fills up. In most designs that's a problem. Here it's a feature — read on.
## 4. "Full" means grown up, not broken
Here's the pivot. When a model that can only add capacity finally fills up, it hasn't failed. **It has
matured.**
Think of the capacity limit as a life stage. A model is *born* as a freshly trained base — its general
education. It has a *working life*, picking up specialised expertise on the job. And it reaches
*maturity* — the point where it has learned about as much as one working life in its niche can teach.
Maturity isn't the end of usefulness; it's the moment the model is most worth learning *from*. So
maturity is the signal to **reproduce**. The capacity ceiling that every other design fights becomes,
here, the clock that times the generations.
Everything now depends on *how* that reproduction happens. This is the heart of the paper.
## 5. Copying rots; combining climbs
Suppose a mature model just teaches a fresh one, and that one teaches the next, and so on down the
line. It's the obvious design — and it fails, for the same reason in AI and in biology.
**In AI terms:** training each generation on the previous generation's output is the exact recipe for
**model collapse**. The model forgets the improbable, loses the rare cases (the tail) first, and drifts
toward its own most common output. And here's the nasty part: the thing that makes teaching-a-student
*useful* — "keep the general, drop the quirky" — *is* the same act that deletes the tail. The operation
you want and the operation that kills the lineage are the same move.
**In biology terms (and for our simplest model it really is the same math — real networks add a measurable twist on top, which we also measure):** picture a model's knowledge as a big bag of
items — facts, skills, behaviours — in certain proportions. One generation is: draw a finite sample
from the parent, and rebuild the child from that sample. That "finite sample" step is *identical* to
**genetic drift** — the way rare gene variants vanish by chance in any finite population. This isn't a
loose analogy; it's the same century-old equations (the WrightFisher model), which is why we can
check our simulations against them exactly. Rare items go extinct first, about ten times faster than
common ones — precisely what drift predicts.
And copying one teacher is **asexual reproduction** — cloning. Biology already knows the fate of a
lineage that only ever clones and never combines: it piles up damage it can never undo, a one-way
decline called **Muller's ratchet**. That's our lens for the *irreversible* part of model collapse —
the capabilities that, once every copy is gone, no amount of combining can rebuild. Naming it that
isn't just poetry — it tells us where to look for remedies, because biology has spent a very long time
solving exactly this.
Two ingredients turn the rot into a climb. Both are things nature does.
**Ingredient one: don't reproduce "dry."** Collapse only happens to a lineage fed *nothing but* its own
output. Mixing in some **real, verified data from the world** — we call this **grounding** — stops it. In
our small experiments, grounding is shockingly cheap: even a few percent of real data keeps most of the
diversity alive indefinitely. But we found an honest limit we didn't expect: grounding can't save the
*very rarest* items at any affordable cost — protecting something of rarity *p* needs a real-data budget
that grows like 1/*p*. Grounding rescues diversity cheaply, but not the deepest tail. Something else has
to do that. That something is sex.
**Ingredient two: reproduce sexually.** Instead of copying one parent, build each new model by
**combining several** — a sexual birth, not an asexual one. AI already has a tool for this: **model
merging**. Why does it help? If several parent models each specialised on *different* things, each one
kept alive rare knowledge the others lost. A combined child inherits the **union** of what its parents
kept — not the tail-thinned *average* of a crowd of near-identical clones. And here's the point that
turns sex from a mere safety net into the engine of the whole thing:
> **A child combined from complementary parents can be *better than any of its parents*.**
Geneticists call this the **FisherMuller effect**: recombination gathers, into one individual,
good variants that arose separately in different lineages — so the child has a combination none of the
parents had. Our simulations show exactly this: combining specialist models that each mastered
different skills produces a model that climbs toward the *best possible* combination — one no single
parent had — while the best single parent, and the plain average of all of them (a "**model soup**"),
both level off well below. This is the paper's core claim in one line: **copying can only recover a
ceiling; combining can break through it.**
And it's not just simulation. In a first test on real language models — three small Qwen models, each
fine-tuned on a different family of tasks, then merged and graded by an automatic checker — **the merge
beat every single specialist**, overall and on every task family. The FisherMuller effect, in real
weights.
That same test pinned down a subtle rule about *how* to combine models:
> **"Merge, don't average" — but only when there's room to lose.**
Keeping each parent whole and **routing** each question to the right specialist beats crudely averaging
them together — *but only when the task is hard enough that averaging actually damages something*. On
easy tasks, a strong model's plain average is already about as good as possible, so the crude soup is
fine. On hard tasks, averaging waters down a hard-won specialist so badly the blend falls below even the
best single parent — and the smarter "keep-them-separate-and-route" approach wins big. So the rule is
precise: **the fancy combining tricks help in exact proportion to how far the plain average is from the
best you could achieve.** A practitioner needs to know this before spending compute on the fancy version.
Two honest caveats, and both are actual findings, not hand-waving:
**Sex can backfire.** When the parents' skills aren't cleanly separable but *tangled together* — when
skill A only pays off if skill B is also present (biologists call this **epistasis**) — blindly
combining two good models can produce a *worse* child, because combining breaks apart a package that
only worked as a whole. Biologists call this **outbreeding depression**, and we reproduce it: on
"tangled" problems, naive merging drops the child below its parents, and the more you mix, the worse it
gets. The design rule: *combine freely when skills are independent; combine sparingly and carefully when
they're tangled.*
**How *widely* you mate matters too.** That last point was about *how much* to mix; a separate knob is
*who mixes with whom*. **Monogamy** = each model only ever combines within a small, fixed circle;
**promiscuity** = any model can combine with any other. Almost all model-merging today is promiscuous by
default — throw everything in one pot. But there's a catch: wide mixing spreads good traits fast, but it
also makes the whole population converge to the *same thing*, killing variety. Narrow, local mixing keeps
separate sub-groups exploring different solutions. We tested this against tangledness, and the best answer
*moves*: on simple (independent-skill) problems, wide promiscuous merging is best; but the more tangled
the skills, the more you want to *narrow* it — full promiscuity converges too fast onto one mediocre
solution and finds a *worse* champion, while keeping structured sub-groups preserves the variety a hard
problem needs. So the rule extends: *merge widely for independent skills; keep separate sub-populations
("island" merging) for tangled ones.*
**AI can do sex better than biology can.** Biology is stuck with two parents, mating more or less at
random, and can't inspect a child before it's born. AI has none of those limits. It can combine **many**
parents at once; it can **choose** which parents to combine, for complementary skills; and it can
**generate many candidate children and keep only the best**, testing them against reality before
committing. We call this **directed sex**, and in our simulations it turns the outbreeding-depression
disaster into a reliable win: where blind combining collapses on tangled problems, directed combining
matches or beats the best parent every time. The real-language-model test shows the same where it can:
breeding many merged offspring and keeping the one that scores highest beats the plain soup on hard
tasks. This is a genuine advantage of *engineered* reproduction over the biological kind, and it's one
of the more useful ideas in the paper.
So §5 in one breath: copying is asexual and rots (Muller's ratchet = model collapse); the cure is to
**ground** every birth in reality and to reproduce **sexually**, combining many complementary parents;
and because AI's version of sex can use many parents, chosen mates, and pre-screened offspring, it's
not just insurance against collapse — it's an engine that makes children better than any parent.
### The limit of sex: models can drift too far apart to merge
Sex has a limit, and it's the sharpest new prediction here. Combining parents works because they're
variations on a shared background. Push two lineages far enough apart and their combination stops being
viable. In biology this is **speciation** — two populations become separate species that can no longer
interbreed. The genetic mechanism has a name (a **BatesonDobzhanskyMuller incompatibility**): a change
that arose in lineage A and a change that arose in lineage B are each harmless on their own, but their
*combination* — which neither lineage ever tested — is broken in the hybrid. A merged model is exactly
such a hybrid. So the theory predicts a specific trajectory as two models drift apart: **they merge
fine → merging starts to hurt → merging becomes useless.**
We built this as an explicit model and confirmed the predicted curve. Three things come out of it:
1. **There's a cliff, and epistasis moves it.** The point where merging fails isn't fixed — it comes
*earlier the more tangled (epistatic) the skills are*. This is a distinct, testable claim:
**at the same amount of drift, whether two models can be merged depends on how tangled their skills
are, not just on how far apart they are.** The existing AI tools for predicting merge success only
measure distance/geometry — they don't have this axis.
2. **It snowballs.** The number of incompatibilities grows with the *square* of the drift, so merge
quality doesn't fade gently — it falls off a cliff. Drift is punished faster than it accumulates.
3. **The design rule:** before merging, weigh how far apart the models are against how tangled the skills
are. Past the cliff, don't merge — **route** instead (keep the specialists separate and pick between
them).
We also did the experiment a skeptic would demand. A known objection: "your 'incompatibility' is just a
loss barrier, and those are famous for being fake — two networks can learn the *same* function but store
it in a shuffled internal order, which *looks* like incompatibility until you line their neurons back up
(a technique called **Git Re-Basin**)." So we tested it directly in real trained networks. We trained
pairs of small networks, merged them, and measured the merge damage *before and after* re-aligning their
neurons. The result is clean:
- Two networks trained on the **same task** but from different random starts: big apparent merge damage,
but re-aligning removes **~98% of it**. That's the fake kind — same skill, shuffled order. (This also
proves our alignment tool works.) And we allowed the aligner *every* legal move for these networks —
not just re-ordering neurons but also re-scaling them — so nothing removable was left on the table.
- Two networks trained on **conflicting tasks**: big merge damage, and even the full aligner removes
**essentially none of it**. That's the *real* kind — genuine incompatibility, not a bookkeeping
artifact. And it can't be waved away as "you just didn't align them well," because the exact same tool
cleaned up the first case. There's even a simple proof that no future alignment trick can fix it: no
single model can obey two rulebooks that contradict each other on the same inputs.
Sweeping from "no conflict" to "total conflict" traces a smooth **incompatibility cliff** in real
weights: the merged model's accuracy slides from 0.97 (no conflict) down to 0.03 (total conflict) — a
hybrid that is literally inviable. So the speciation effect is real, not a relabelled artifact.
And we ran the honest flip side, deciding in advance to report it either way: what if two networks just
*specialise differently*, with no conflict at all — one keeps training only on digits 04, the other
only on 59? Do they drift into incompatibility on their own? **No.** At every amount of divergence we
tested, the merge damage stayed at zero — and the merged model actually *rescued* the two specialists:
each parent alone had forgotten half the digits (scoring ~0.50), while their merge scored ~0.95. So in
these experiments, models don't become unmergeable just by growing apart; they become unmergeable when
they learn things that genuinely *contradict*. That's good news for merging — specialisation is safe,
conflict is the danger — and it makes the theory's prediction sharper, not weaker.
One question is left hanging, and the rest of the paper is about it: combining preserves *what the
parents kept* — but **who decides what each parent keeps, and which children are worth keeping?**
## 6. Don't design the selector — evolve it
There are two ways to answer that question, and the first one is wrong. We could try to hand-write the
rule for "what knowledge to keep and pass on." But nobody actually knows that rule. "Keep the general,
drop the specific" sounds wise until you ask *which* generalisations, in *which* domain, at *which* level
of detail — and it falls apart. This is the deepest gap in the whole scheme, and you can't fill it by
decree.
The second answer is the one nature used: **don't design the selector — let it evolve.** Let different
models carry different *policies* about what's worth keeping and combining. Let the policies that produce
strong children spread, and the policies that produce weak children die out with their lineages. What the
lineage considers *important* — its "taste" — is discovered by selection, not imposed by us.
So **two things get inherited, on two channels.** The *content* — the actual knowledge — is passed down
directly (this is the "Lamarckian" part: inheriting things acquired during a lifetime, which biology
forbids for genes but culture allows for ideas). The *selection policy* — what to keep, who to breed
with, which children to screen for — is *itself* inherited, varies between models, and survives in
proportion to how well it works. That second channel is Darwinian. The system is both at once: it
inherits *content* like culture, and selects *policies* like evolution.
And it closes neatly: Darwinian selection needs a pressure — something that decides which policies win.
That pressure is already in the design. What tells a lineage its taste was good? Whether its children
succeed *against reality*. The reality check that stops collapse (grounding, §5) and the fitness signal
that guides the evolving taste turn out to be the **same thing**, seen from two angles.
## 7. The big danger: what you measure is not what you want
Adding selection adds selection's classic trap, and it's dangerous enough to sink everything if ignored.
Evolution optimises, ruthlessly and without foresight, for exactly what you *measure* — never for what
you *meant*. (In ML you know this as reward hacking or specification gaming; economists call it
Goodhart's law.) Get the fitness measure a little wrong and the lineage will exploit the gap more
cleverly than any rule you could write.
For a *knowledge* lineage there's a particularly nasty version. The natural way to measure how "good" an
idea is might be *how well it spreads* — but a false-but-persuasive idea spreads beautifully. Human
culture is full of highly contagious nonsense; confident wrongness routinely beats careful accuracy. Let
selection loose on models without care and it will breed a lineage that is fluent, compelling, and
**wrong** — model collapse with an optimiser actively steering toward the cliff.
Only one thing makes fitness track *truth* instead of *appeal*: **being judged against a reality that
can say no.** Fitness has to mean "did this knowledge correctly predict what the world actually does when
you act on it" — not approval, not fluency, not a gameable benchmark score. This is why the reality check
matters twice: it's both the anchor that stops passive collapse *and* the only thing that keeps the
evolving taste honest.
The second danger is **everyone converging to the same thing**, and avoiding it takes work at two levels,
because selection can only preserve variety that already exists — variety first has to be *supplied* and
then *kept*.
- **Supply.** A lineage that only learns from the accredited elite has a monoculture for a source — the
"top" experts are, by definition, the ones who won the consensus. So the system must deliberately learn
from **outliers and heretics** too — not out of fairness, but because diverse starting material is the
raw fuel for everything downstream.
- **Preserve.** Even with varied input, plain "keep the best" selection converges — it stampedes toward
the single current champion and wipes out the rare specialists. The fix is well known:
**quality-diversity** selection, which rewards being *good* **and** being *different* at the same time,
keeping complementary specialists alive. In our simulations this is decisive: greedy "keep-the-best"
collapses diversity almost immediately and gets stuck on a mediocre answer, while quality-diversity
keeps the very specialists that sexual combining then needs as parents.
The two levels meet at reproduction: multi-parent combining is the *vehicle* that carries the preserved
diversity into the next generation. Supply the variety, preserve it, recombine it — remove any one and
the lineage collapses onto its own first guess.
## 8. A society needs institutions, not just experts
One requirement is easy to miss and fatal to skip. The easy part of a society is having specialists. The
*hard* part — which human civilisation took millennia to build — is the **institutions that let fallible
specialists combine without each re-checking everything**: reputation, replication, credentials, and
above all **peer review**. These are error-correction systems, and they exist because a group of
unreliable experts left to reinforce each other is *more* wrong than any one of them alone.
This is exactly where today's multi-agent AI fails: put several models in a room and they tend to agree
sycophantically and confabulate together, because they have all the specialisation and none of the
institutions. A real multigenerational society has to specify not just how models learn, reproduce, and
get selected, but how they *check each other* — how a wrong model loses standing *before* its error gets
merged into its children and inherited.
## 9. The lineage must stay open to reality
However many generations deep it goes, a society of models shares one hard limit: it has only ever
*read*. Its entire inheritance is a record of things that were *said*. It lives on the bottom rung of
what causality researchers call the ladder — **observation** — and no amount of reading ever reaches
**intervention**. Watching doesn't tell you what would happen if you *acted*; correlation isn't causation
at any scale.
Only intervention — actually reaching out and changing the world to see what happens — climbs that
ladder, and a language model can't intervene. That's what humans and their instruments provide, and the
gift isn't "truth," it's **constraint**: reality's unique power is that it can say **no**. Text just
offers more opinion; an experiment delivers a refusal no consensus can overturn. As before, that refusal
does double duty: it stops collapse *and* keeps the evolving taste selecting for truth over persuasion.
Two honest riders. First, the human reality signal is *dirty* — warped by publication bias, incentives,
and the occasional fraud — which is exactly why the error-correcting institutions of §8 have to sit at
the humanmachine boundary. Second, humans are the *current* supplier of intervention, but the hands-on
half is being automated (robot labs already run their own experiments). What looks durable in the human
role isn't the hands — it's the **choice of what to test and which refusals matter**.
## 10. Why this is affordable
A practical fact turns this from daydream into buildable proposal: **it almost never re-pays for the one
genuinely expensive thing in AI — training a model from scratch.**
Training a foundation model from scratch costs a fortune in data and compute. This design does none of
that per generation. Every model is *born* from an existing open model that already paid that cost;
specialising one is a small patch trained in hours on a single consumer GPU; running the society is
ordinary use; and reproducing — merging parents into a child — can be done directly on the weights with
*no retraining at all*. Selection costs more (you have to run populations and discard the losers), but
that's a multiplier on an already-cheap unit, not on a from-scratch budget.
The economics only work with **open-weight** models — for practical and legal reasons at once. You have
to be free to inspect, modify, and redistribute the weights, and most proprietary licences forbid using
a model's output to train another (which is exactly what reproduction here does). That's not ideology —
it's a structural constraint, and a democratising one: it puts the whole thing within reach of a single
lab.
## 11. Can it grow forever? Baking knowledge back into the base
One thing we quietly assumed: can the lineage accumulate *without end*? The individual model is
bounded — that's the clock. But the lineage looked unbounded, each generation starting a step ahead.
Look closer and a second budget also fills up.
Every new model is a clean base plus an inherited pile of *soft* patches — the acquired knowledge carried
in add-on modules rather than baked into the core. Those patches are what make the lineage
multigenerational — but they're not free: they slow the model down, and past some depth they're better
*consolidated* than carried. The lineage matures too.
The fix is the same operation, one level up. When a lineage's acquired knowledge has proven stable across
enough generations, **re-mint the base**: bake all those accumulated patches into the *weights* of a
fresh from-scratch-scale model — a new base that is *born already knowing* what took many generations to
learn in patches. The soft budget resets; the next era starts from a higher floor. What was hard-won and
*learned* becomes cheap and *built-in*. (Biologists have a name for acquired traits that eventually
become innate: the **Baldwin effect**.)
Three honest riders, because this is the most consequential step:
- **Cost.** This is the one step that re-pays part of the from-scratch bill — the exception to §10. It's
bearable only because it's *rare*, spread thin over many cheap generations.
- **Irreversibility.** Until now, one thing was always recoverable — the original clean base, whose lost
rare knowledge you could restore just by reloading the file. Bake the current lineage into new
permanent weights and that escape hatch closes: if the lineage had been quietly collapsing, re-minting
*freezes the collapse in place* and throws away the one uncollapsed reference that could have caught it.
In our experiments this happens exactly as feared — and a cheap safeguard prevents it: **only re-mint
while the lineage is provably healthy and diverse**, never as a rescue for one already drifting.
- **Branching.** Different labs will re-mint on different criteria and produce different bases; the
lineage branches. That's not a bug — it's the tree of life, and it's exactly what open weights make
possible.
So can it grow forever? **Yes — but only because it forgets and consolidates at every level, including
the base.** Nothing is stored without limit anywhere. Unbounded growth of *capability* is bought with
*bounded* storage plus periodic consolidation.
## 12. One idea at four speeds
Step back and it all resolves into a single process running at four nested speeds. The **downward**
motion is transmission — passing hard-won knowledge down:
1. **Within one model, over its working life:** experience moves from fast memory into slow, durable
weights, without catastrophic forgetting.
2. **Between generations, at maturity:** mature models reproduce, recombined into a fresh one.
3. **Across many generations:** each inherits the compressed achievements of the last and adds a little.
4. **Across eras:** a proven lineage's accumulated patches get baked into a re-minted base, becoming
innate.
The first and last are the *same operation at opposite ends of the scale* — fast/soft memory
consolidating into slow/hard memory — one running overnight inside a single model, the other across an
era inside a whole society. The **sideways** motion is selection — acting across the population at every
speed, on the policies that decide what gets passed on, with reality as the judge and diversity kept
alive so the specialists survive.
Three rules govern all of it: **reproduce by combining, not copying, or you rot; keep the disagreements
and surprises, or you converge; and anchor fitness to a reality that can say no, or you drift toward
whatever is merely convincing.**
## 13. What we actually built, found, and left open
We didn't just argue this — we built small, fully reproducible models to test it, plus a first bridge to
real language models. Here's the shape of the evidence (a separate document has the numbers).
**What we built and found:**
- **An exact account of collapse.** Because generational training *is* the genetic-drift process, we can
check our simulator against century-old formulas — and it matches to a fraction of a percent. Collapse
isn't argued by analogy; it's derived.
- **Grounding is cheap — but has a limit.** A few percent of real data keeps most of a lineage's
diversity alive indefinitely — but not the very deepest rare cases, which need combining. This is what
makes a continually-learning society economically realistic rather than a data-hungry fantasy.
- **"Merge, don't average."** Combining several teachers by *averaging* their outputs — the obvious
thing, and what a "model soup" does — mathematically cancels the benefit of having several teachers. A
*merge* that keeps each item's strongest source realises it. Most current multi-model setups get this
wrong by default.
- **Collapse and its cure in real networks, and on real images.** We reproduced the same effects in small
neural networks and in a generator of handwritten digits (MNIST): a model trained on its own output
collapses to a single blurry digit, while a little grounding keeps all the styles alive. Honest
wrinkle: real networks *smooth* things over, so the naive diversity metric can mislead — the right
measure is distance-from-truth.
- **Sex that beats the parents — and when it doesn't.** In simulations, combining complementary
specialist models produces a model better than any parent (FisherMuller), climbing toward the best
possible combination as you add more, more-diverse parents — while averaging and best-single-parent
level off below. On *tangled* problems, blind combining instead produces below-parent children
(outbreeding depression) — and *directed* combining (choose mates, screen offspring, many parents)
reliably fixes it.
- **Monogamy vs promiscuity.** Sweeping how *widely* models merge — from local/monogamous to
everyone-with-everyone/promiscuous — against how tangled the skills are, the best breadth **shrinks as
the skills get more tangled**: wide promiscuous merging wins when skills are independent, but on tangled
problems it converges too fast onto one mediocre solution and finds a worse champion, so keeping
structured sub-populations wins. (Promiscuity always lifts the *typical* model but always destroys
variety.) A merging design knob the field, which throws everything in one pot, doesn't currently have.
- **The combining claims, in real language models — with a sharp condition.** Merging fine-tuned Qwen
models (up to 7B on a GPU cluster) produces a generalist that beats every specialist parent; and
keeping parents separate and *routing*, or *breeding and screening* offspring, beats the plain average
*but only when the task leaves room to lose*. On easy tasks the plain average is already at the
ceiling and the tricks add nothing; on hard tasks the average dilutes a specialist below even the best
single parent, and the smarter operators win clearly. This is a prototype (three task families, one
run), so we read it as *signs, not exact numbers* — the *whole grounded society* on a language model is
still the open step.
- **The whole society, and why every part is needed.** In a population evolving on a "reality" landscape,
the full system — grounding + combining + preserved diversity — climbs to the top while keeping its
specialists. Remove *grounding* and it collapses into a confident, wrong consensus (a direct analogue
of training on the internet's growing pile of AI-generated text); remove *combining* and it gets stuck;
remove *diversity* and it converges too fast to a worse answer. Each removal breaks differently; only
the whole thing climbs.
**What's borrowed vs. what's ours.** We're careful about credit because the area is crowded. **Already
known (and we claim none of it):** that model collapse is genetic drift; that a merged model can beat its
parents; that decorrelated parents merge better; that naive averaging is worse than smarter merges; that
populations of self-improving models can climb. **What's genuinely new here** is the *theory* those
results have outrun — a real population-genetics of sex for model societies, which *predicts* rather than
just *observes*: the "merge, don't average" law, FisherMuller as the reason children beat parents,
outbreeding depression on tangled landscapes (turning "when does merging help?" into something the
problem's structure predicts), grounding as a migration-vs-drift balance with a critical real-data
fraction, directed sex as AI's distinct advantage, the whole integrated society whose parts are shown
*jointly necessary* — and, newest of all, **model speciation**: the account of *when two models are too
far apart to merge at all*, confirmed in real weights.
**What's still open — honestly.** We fill the old hole (what to select) by *evolving* the selector
instead of designing it — but the hole *moved* rather than closed, and the new one is harder: **the
fitness function.** What reality-anchored measure selects for *truth* without also selecting for
*persuasion*, given that in our own species the two have been at war for the entire history of ideas?
Alongside it: the *institutions* that let models correct each other before error is inherited (§8), which
we don't solve; and the *calibration* of all the knobs the experiments left open — how many parents, how
complementary, at what ratio of inherited-to-real data, how healthy a lineage must be before its
knowledge is safe to make permanent. These are at least *measurable*, which is the difference between an
open problem and a hole. And the biggest gap: the *combining* claims now hold in real language models,
but the *society* — the grounded, diversity-preserving, continually reproducing loop — does not yet. The
real test is to build that whole thing out of actual open language models and see whether all the signs
survive. **The operators, checked; the living society, next.**
---
*For the full argument, the literature it's positioned against, the exact predictions, and the
references, see the complete paper: `the-evolution-of-sex-for-ai.md`.*

View file

@ -1,919 +0,0 @@
# The Evolution of Sex for Artificial Intelligence
### A population-genetic framework for societies of agents that reproduce, recombine, and stay open-ended
*A perspective, written from a geneticist's chair. Companion to a set of minimal, reproducible working
models and a first language-model prototype (both built).*
**Giorgio F. Gilestro** · Department of Life Sciences, Imperial College London ·
giorgio@gilest.ro · https://lab.gilest.ro
---
### A note on vocabulary (please read this first)
This paper sits at the meeting point of three fields, and it is written so that a reader from any
one of them can follow all of it. We therefore **spell out** each field's jargon the first time it
appears, even at the risk of belabouring the obvious for the specialist. A short glossary, in case
you skipped a definition:
- **Model collapse** *(machine learning)* — the degeneration that happens when you train a model on
data produced by earlier models, over and over: rare cases disappear and the model drifts toward a
bland average.
- **Distillation** *(machine learning)* — training a fresh "student" model on the outputs of one or
more "teacher" models, so the student ends up knowing a compressed version of what they knew.
- **Model merging** *(machine learning)* — combining several trained models directly, at the level
of their weights, into one — no retraining. (Think of it as breeding two models rather than
teaching a third.)
- **Genetic drift** *(population genetics)* — the random loss of rare variants that happens in any
finite population simply because not everyone leaves offspring. It is the neutral, no-selection
baseline of evolution.
- **WrightFisher process** *(population genetics)* — the standard mathematical model of drift. Our
minimal model of knowledge transmission *is* this process exactly; a real trained network is this
process plus a measurable, architecture-specific bias we quantify.
- **Recombination / sexual reproduction** *(biology)* — making an offspring by combining pieces from
more than one parent, rather than copying a single parent (which is *asexual* reproduction).
- **Muller's ratchet** *(population genetics)* — the way an asexual lineage, one that never
recombines, accumulates damage it can never undo. We will argue it is the right lens for the
*irreversible* part of model collapse — the capabilities that, once lost from every parent, no
merging can rebuild.
- **Catastrophic forgetting** *(machine learning / neuroscience)* — a neural network overwriting what
it knew when it learns something new.
We have tried to keep the big picture legible on every page, and to be candid about what is argument
and what is evidence. The evidence is mostly from **deliberately small models** — mathematics, small
neural networks, image generators, and evolutionary simulations. A first bridge to real language
models exists — a prototype that recombines LoRA-specialised Qwen models up to 7B on a GPU cluster,
which confirms the recombination signs (below) — but the *full grounded society* has not yet been
built on a large language model. We will say so repeatedly, because the gap matters.
---
## Abstract
AI is turning from single frozen models to **populations of agents** that persist, specialise, and are
increasingly *recombined* into new models — a shift visible in multi-agent societies, population-based
self-improvement, and the explosion of **model merging**. The field is doing this with the vocabulary
of evolution — "crossover," "mutation," "mate choice," "offspring that beat their parents" — but as
loose metaphor draped over search algorithms. This paper argues that a rich, quantitative body of
applicable theory already exists in the branch of biology that studies exactly this: the **evolution
of sex**. Ninety years of population genetics analyse when reproducing a population by *recombination*
beats copying, when it backfires, and how to do it better — and, read as an engineering framework, it
supplies overlooked variables and testable design rules for keeping a society of models learning
across generations instead of decaying. The underlying shift of perspective is the contribution we
most want to land: **treat multigenerational model populations as systems whose inheritance,
diversity, and compatibility must be managed — not merely as collections of models to optimise.**
One diagnosis anchors the frame: training each generation on the last is **genetic drift**, and the
resulting **model collapse** is the loss of rare variants a finite population always suffers (the
WrightFisher process). We reached that account independently; it has also been formalised in
parallel by others (Shumailov et al., 2024; Riis, 2026), whom we cite for priority of publication —
a convergence we read as corroboration of the frame. This paper is about the structure the diagnosis
opens: the remedy side and its limits. Single-
teacher copying is **asexual** reproduction, and the irreversible arm of its decay corresponds to
**Muller's ratchet** (a correspondence we state with its scope, not as identity); the remedy biology
found for the ratchet is **sex**. A society of models should reproduce sexually — each new model
**recombined from several complementary parents** (which the field already does, as *model merging*),
selection **anchored to a reality that can say no** (not to the consensus of other models), and
diversity actively **preserved**. In our models — from closed-form to trained networks to a
language-model prototype — those three ingredients together let a lineage not merely avoid collapse
but **climb**, producing models fitter than any ancestor (the **FisherMuller effect**) while each
specialty is re-earned and exceeded; whether the full recipe holds at frontier scale is the open
question the framework is built to test.
From the geneticist's apparatus we extract falsifiable, load-bearing claims (each stated with its
operator and scope in the text): (i) **"merge, don't average"** — a conservation result: refitting a
child to the *mean of its parents' output distributions* conserves expected rare-capability mass at
the single-parent level, cancelling the multi-parent gain *to first order in the rare-item regime*
(outside it, variance reduction from averaging can help — the result is a first-order cancellation,
not a universal impossibility), while union-preserving operators realise the gain in all regimes —
derived in the minimal model, with its weight-space image the headroom rule below; (ii)
**offspring can exceed every parent** (FisherMuller), the real argument for sex in model societies;
(iii) on **rugged, epistatic** task landscapes, blind recombination causes **outbreeding depression**,
yielding a design rule — *merge freely when skills are additive, sparingly and with selection when
entangled, and route rather than blend under overlap*; (iv) **grounding is immigration** from a
non-drifting reality, giving a critical real-data fraction far below one; and (v) — the sharpest new
prediction — sex has a **limit**: as two models diverge they undergo **speciation**, a
merge-compatibility cliff (compatible → outbreeding depression → hybrid inviability) whose onset is set
by divergence *and* epistasis via **BatesonDobzhanskyMuller incompatibilities**, and whose damage
grows *super-linearly* (the OrrTurelli snowball). We introduce and model this "model speciation"
directly, and confirm it in real trained weights: a merge barrier that survives alignment under the
*full* function-preserving symmetry group of the network (not just Git Re-Basin permutations), rising
with functional conflict while hybrid fitness falls to inviability — with an honest converse we
pre-registered and found: absent conflicting training signals, divergently-specialised lineages of
shared ancestry developed *no* isolation at any divergence tested, the merge instead *rescuing* the
forgetting specialists. Isolation must be provoked by conflict; specialisation alone did not speciate.
AI also has an advantage biology lacks: **directed sex** — unbounded parents, chosen mates,
and offspring screened before they are kept — engineered recombination with a flexibility of parent
choice and pre-deployment screening that natural mating systems do not approach.
We support the argument with **minimal, reproducible models** — a closed-form-exact account of drift
and grounding, the same effects in small trained networks and an MNIST image generator, a real-weight
demonstration of the speciation cliff (a Git Re-Basin residual that survives neuron alignment), and
evolutionary simulations of the whole society — and a first **language-model prototype**: merging
LoRA-specialised Qwen models (to 7B on a GPU cluster) yields a generalist that beats every specialist
parent, with the sharp headroom condition under which "merge, don't average" bites. The scope is
honest: these are existence proofs and design rules; the *whole grounded society* on a large language
model is the open step. We position the work carefully against the crowded 20252026 landscape of
evolutionary-AI and merging methods — conceding what they own and marking, precisely, what a genuine
population-genetics of sex adds.
---
## 1. From a society in space to a society in time
The idea of many AI agents working together — a "society of mind" (Minsky, 1986), or today's
multi-agent systems — arranges intelligence across *space*: several specialists side by side,
dividing a task. This paper is about a different axis: *time*. Not a society that merely exists at
one moment, but one that **persists and renews across generations**, each new cohort of models
starting from the compressed knowledge of the last.
The unit that matters is therefore the **generation**, and the event that matters is **reproduction**:
the making of a new model from older ones. A single model, like a single mind, is bounded and
eventually stops improving. A *lineage* need not be. Human civilisation is not clever because any one
person is; it is clever because each generation inherits the distilled achievements of the previous
one and adds a little. We propose building AI the same way — and, crucially, getting the *reproduction*
right, because that is exactly where it can go wrong.
### Where this sits, and what is new
This axis is suddenly crowded. By 2026 several groups build **populations of models or agents that
improve across generations**: societies of independently-specialised models that self-improve for more
rounds than a single agent (Multiagent Finetuning — Subramaniam et al., 2025); open-ended archives of
self-rewriting coding agents (the DarwinGödel Machine — Zhang et al., 2025); groups that evolve by
sharing experience across branches (Weng et al., 2026); persistent agent *ecologies* with reproduction
and cumulative culture (TerraLingua — 2026). In parallel, **model merging** has become a small industry
with an overtly evolutionary vocabulary: crossover-mutation-selection over LLM populations (GENOME —
2025), niching and "mate choice" (Sakana's M2N2 — 2025), and evolutionary search over merge recipes
(Akiba et al., *Nature Mach. Intell.* 2024/25).
We are candid about the consequence. Three things we do **not** claim. First, that collapse is
WrightFisher drift: formalised independently (Riis, 2026; Shumailov et al., 2024), sharpened to a
closed-form first-extinction law whose onset coincides with collapse (Benati et al., 2025) and to a
quantitative-trait-genetics account for diffusion models (Yoon et al., ICLR 2025), and conceded here.
Second, the bare empirical facts that a merged model can beat its parents, that decorrelated parents
merge better, and that naive averaging is inferior to sign- or routing-based merges (TIES, DARE,
mixture-of-experts routing): all established. Third, that merge success can be *predicted at all*:
machine-learning-native predictors exist, from interpretable pairwise metrics (gradient alignment —
Zhou et al., 2026) to capacity/rate-distortion accounts of "merging collapse" (2026); what they lack,
and we supply, is the *mechanism* — when and why the failure is a coordinate artefact versus genuine
functional incompatibility, and what moves the cliff. What a geneticist is placed to supply is a
**framework** rather than a search heuristic. The nearest precursor is a theory-of-computation
tradition reading sex as an algorithm for *mixability* (Livnat & Papadimitriou, 2016), pre-dating
model merging; the works above use evolution chiefly as vocabulary over an optimiser, and — to our
knowledge — the quantitative apparatus of the evolution of sex (FisherMuller, outbreeding depression,
migrationdrift balance, reproductive isolation) has not previously been carried over as more than
metaphor. We are also candid about what *kind* of contribution each of our claims is, because three
different things are easily conflated: **interpretation** (an existing result is usefully understood
in these terms — e.g., merged offspring beating their parents as FisherMuller), **explanation** (the
transferred mechanism accounts for observations existing accounts leave open — e.g., which merge
failures are coordinate artefacts and which are functional), and **prediction** (the framework
forecasts an unmeasured outcome and improves a design decision). This paper is strongest on the
first, makes concrete progress on the second, and reports a first, bounded step on the third: a
**controlled predictive test** at small scale in which pre-merge *functional-disagreement* measures —
chosen by the framework — showed a detectable, held-out-robust association with merge damage on a
constructed task grid, while the selected weight-geometry baselines did not. We are precise about
that result's boundary where it is reported: it is a small-model demonstration on a constructed grid;
the proposed epistasis-specific refinement did not outperform plain disagreement; predictor
differences are not individually significant head-to-head; and whether the prediction improves a
budget-matched operator choice remains open. The organising shift we argue for
is prior to any single mechanism: **treat multigenerational model populations as systems whose
inheritance, diversity, and compatibility must be managed — not merely as collections of models to
optimise.**
## 2. Why today's models cannot do this
Today's large language models have no life cycle. They are trained once, at enormous cost, then
**frozen** and deployed as a fixed artefact that does not learn from the people it serves. Learning
and doing are split into two eras with no bridge between them.
There is a real reason for the freeze. Updating a neural network on new information tends to overwrite
what it already knew — **catastrophic forgetting**, a problem understood since the late 1980s
(McCloskey & Cohen, 1989; French, 1999). Freezing avoids it by refusing to learn at all. The result
is a mind with no childhood, no growth, and no way to pass anything on. A lineage needs the opposite:
members that learn through their working lives, reach maturity, and hand on what they gained. So the
first requirement is a learner that can grow *safely*.
## 3. A learner that can grow without forgetting
The individual model needs two properties.
**It must not catastrophically forget.** Instead of overwriting its core as it learns, it keeps that
core frozen and only *readable*, and carves each new skill into freshly-added capacity beside it. In
machine learning this is called *parameter isolation* (progressive networks — Rusu et al., 2016;
prune-and-freeze — Mallya & Lazebnik, 2018; and, most practically, **LoRA** and other small trainable
"patches" bolted onto a frozen model — Hu et al., 2021). If the core is never altered, its *parameters*
cannot be forgotten — though a precise reader should note the system's *behaviour* can still shift
while adapters are active, so the guarantee is of a recoverable core, not of unchanging conduct. This
is what lets a model accumulate a coherent working life of expertise — the kind of stable knowledge
worth passing on.
The brain offers a partial blueprint. *Complementary Learning Systems* theory (McClelland,
McNaughton & O'Reilly, 1995) — itself a response to the forgetting problem — describes two subsystems:
a **fast** store (the hippocampus) that grabs an experience in one shot, and a **slow** store (the
neocortex) that integrates regularities gradually without disruption. We do not lean on any particular
account of how the brain moves knowledge between them; the architecture needs only that *some*
periodic **offline consolidation** step exists, moving knowledge from the fast store to the slow one
when the system is idle. The machine version is clean regardless: the prompt is working memory, an
external database is the fast episodic store, the trained weights are the slow store, and consolidation
migrates the first into the last.
**It is bounded.** Because the model only ever *adds* capacity and freezes what it has, it eventually
fills up. In most designs that is a wall to dread. In ours it is a clock.
## 4. "Full" is maturity, not failure
Here is the pivot. A bounded learner that fills up has not broken. **It has grown up.**
Read the capacity limit as a life stage. A model is *born* as a freshly-schooled base — its general
education. It enters a **working life**, adding specialised knowledge as it does its job. And it
reaches **maturity**: the point where it has learned much of what one working life in its niche can
teach. Maturity is not the end of usefulness — it is the moment the model is most worth learning
*from*. So maturity is the cue to **reproduce**. The capacity ceiling that every other architecture
fights becomes, in ours, the metronome of the generations.
Everything now turns on how that reproduction is done — and this is where the paper's central claim
lives.
## 5. Reproduction: copying collapses, recombination climbs
Suppose a mature model simply teaches a fresh one — distillation, one teacher to one pupil, generation
after generation. This is the obvious design, and it fails, for a reason that is exactly the same in
machine learning and in biology.
**The machine-learning statement.** Training each generation on the previous generation's outputs is
the recipe for **model collapse**: the model forgets the improbable, loses the *tail* of the
distribution (the rare cases) first, and drifts toward its own most common output (Shumailov et al.,
2024). Worse for us, the very rule that makes distillation useful — *keep the general, drop the
idiosyncratic* — **is** tail-deletion by design. The operation that would power a cultural ratchet and
the operation that drives model collapse are the same act.
**The population-genetics statement (the same thing, for the minimal model).** Represent a model's
knowledge as a distribution over discrete "items" — capabilities, facts, modes of behaviour. One
generation is: *draw a finite sample from the parent, and refit the child to it.* In this **minimal
inheritance model** the finite-sampling step is **exactly** genetic drift — the random loss of rare
variants in a finite population — described by the century-old **WrightFisher** model (Wright, 1931;
Fisher, 1930): the same equations, which we use as closed-form checks on our simulations. Rare items
go extinct first, roughly ten times faster than common ones, precisely as drift predicts. **The
boundary of the identity matters, and we measured it:** real neural training adds approximation,
optimisation noise, and inductive bias on top of sampling, and when we fit trained networks against
the exact drift null they deviate in *opposite, architecture-specific directions* — a smoothing
recurrent model resists collapse (it keeps spurious variants alive), a sharpening image generator
accelerates it (our learning-kernel result, below). So the honest statement is: the minimal
inheritance model is exactly WrightFisher; a real learner is WrightFisher *plus a signed,
measurable estimator-bias operator* — and the drift signs (rare-first loss, the grounding response)
survive that operator in every architecture we tested.
And single-teacher copying is **asexual reproduction** — cloning one parent. Nature already knows what
happens to an asexual lineage that never recombines: it accumulates damage it can never repair, a
one-way decline geneticists call **Muller's ratchet** (Muller, 1964). We use the ratchet as the
*organising correspondence* for model collapse, with its scope stated: strictly, the ratchet is the
stochastic loss of the least-degraded class under recurring deleterious change in an asexual
population, so it maps onto the *irreversible* component of capability loss (once every copy of a rare
capability is gone from all parents and sources, no recombination can rebuild it) rather than onto
every form of degradation. That is exactly why the correspondence is useful rather than decorative: it
says the cure must act *before* fixation-by-loss — keep complementary variants alive somewhere in the
population — because recombination can only reassemble what still survives. Biology solved this
problem, and its solution is the subject of this paper.
Two ingredients turn the collapse operation into a climb. Both are things nature does.
**First: do not reproduce "dry."** Model collapse is a property of a lineage fed *only* its own
output; the documented fix is that keeping some real data in the mixture arrests it (Shumailov et al.,
2024). We call that real data **grounding** — fresh contact with the world, verified against it. In
our minimal models, grounding is startlingly cheap: mixing in even a few percent of verified real data
holds on to most of the diversity indefinitely. But — an honest limit we found and did not expect —
grounding cannot save the *very rarest* items at any affordable budget; protecting an item of rarity
*p* needs a real-data budget that grows like 1/*p*. Grounding rescues diversity cheaply; it does not,
by itself, rescue the deep tail. Something else must. That something is sex.
**Second: reproduce sexually.** Instead of copying one parent, build each new model by **recombining
several** — a *sexual* rather than asexual birth. In machine learning this already has a name and a
working implementation: **model merging** (Akiba et al., 2024). Its importance here is not efficiency;
it is that recombination does something copying cannot. If several parent models have each specialised
on different parts of reality, each has kept alive rare knowledge the others lost. A recombined child
inherits the **union** of what its parents kept — not the tail-thinned *average* of a crowd of
near-identical copies. And here is the point that lifts sex from a safeguard to the engine of the whole
scheme, and the reason biology invented it:
> **An offspring recombined from complementary parents can be *fitter than any of its parents*.**
Geneticists call this the **FisherMuller effect** (Fisher, 1930; Muller, 1932): recombination brings
together, in one individual, beneficial variants that arose separately in different lineages, so the
child holds a combination none of the parents had. In our simulations this is exactly what we see —
recombining decorrelated specialist models yields a model that climbs toward the best-possible
combination, a genotype *no single parent possessed*, while the best single parent, and the naive
average of all of them (what the field calls a "model soup" — Wortsman et al., 2022), both plateau
well below. This is the concrete meaning of the paper's title claim, "the lineage climbs in general
knowledge; specialisation is re-earned each generation," and it is why the reframing from
teacher→pupil to *sexual reproduction* is not cosmetic: **copying can only recover a ceiling;
recombination can exceed it.**
This is no longer only a simulation. In a first language-model prototype — LoRA specialists on
disjoint task families, recombined and judged by an exact verifier — a merge of three specialist Qwen
models (7B, on a GPU cluster) **beats every single specialist**, overall and on every family: the
FisherMuller effect, in real weights. The same prototype pins down *when* the finer "inherit the
union, don't average" rule actually bites. Keeping each parent whole and **routing** each input to the
right one beats the tail-thinning average — but only when the task is hard enough to leave room to
lose: on easy tasks a strong model's plain average is already at the ceiling, so the crude soup is
fine, whereas on hard tasks the average dilutes a hard-won specialist so badly it falls below even the
best single parent, and routing wins by a wide margin. The rule is therefore precise: **the union
beats the average in exact proportion to how far the average is from the best attainable** — a caveat
that sharpens rather than weakens the claim, and that a practitioner needs before spending compute on
the fancier operator.
**The operator boundaries (stated, because "merge, don't average" is not one claim but a family).**
Four different operators travel under these words, and the conservation result belongs to exactly one
of them. What is *derived* is this: when a pupil's knowledge is refit to the **mean of the parents'
output distributions**, the expected mass on any rare item is conserved at the single-parent level —
in the rare-item regime (`n·p/K ≪ 1`) the 1/K dilution of averaging cancels the union gain of having
K parents to first order — outside that regime, survival is convex in mixed mass and averaging's
variance reduction can help, so this is a first-order cancellation, not a universal impossibility;
whereas an operator that keeps, per item, its **strongest source** (and renormalises, which itself
redistributes mass) realises the union in all regimes. That statement is exact in the minimal model, and it presupposes an
oracle (or verifier) able to say which source is strongest. The two operators the LLM prototype
tests — **weight averaging** (a nonlinear network's weight-mean does not compute the mean of its
parents' outputs) and **routing among intact specialists** (which keeps K models' storage and an input
classifier, a different parameter and inference budget from one fixed-size child) — are *empirical
cousins* of the two sides of that law, not instances of it. The headroom rule above is precisely the
empirical bridge: it says when the weight-average behaves like the diluting mean (hard tasks, weak
base) and when a capable base absorbs the dilution (easy tasks). And all of it operates within a
capacity boundary: when parental capabilities genuinely cannot coexist in the child's capacity, no
operator preserves the union — that regime is the subject of the speciation section below.
Three results keep this honest, and all are results, not hand-waving.
*Sex can backfire.* When the parents' skills are not cleanly separable but **entangled** — when the
value of one capability depends on which others are present (geneticists call this **epistasis**) —
blindly recombining two good models can produce a *worse* child, because recombination breaks up a
combination that only worked as a whole. Biologists call this **outbreeding depression**, and we
reproduce it: on "rugged" (highly entangled) problems, naive merging drops offspring below their
parents, and the more you mix the worse it gets. The design rule that falls out is simple: *merge
freely when skills are complementary; merge sparingly, and carefully, when they are entangled.*
*The mating system matters too — not just who mates, but how widely.* The result above is about the
recombination *rate*; a separate knob is the population's **mating structure** — whether reproduction is
**monogamous** (each model recombines within a narrow, local circle) or **promiscuous** (mates drawn
freely from the whole population). Almost all model-merging implicitly assumes promiscuity — fuse
everything, or route over one flat pool — but population genetics says the breadth of gene flow is itself
consequential, because wide flow spreads good variants fast while **homogenising** the population, and
narrow flow preserves the distinct sub-populations needed to explore several solutions at once (Wright's
*shifting balance*). We sweep exactly this breadth against landscape ruggedness, and the optimum moves:
on smooth (additive) landscapes wide, promiscuous mating is best (spread the one good direction fastest),
but as the landscape gets rugged the best breadth **shrinks to an intermediate value** — full promiscuity
prematurely converges onto one basin and finds a *worse* champion, while pure monogamy over-fragments.
Throughout, wide mating lifts the *typical* model but monotonically **destroys diversity** — so on rugged
problems, where the best model needs preserved diversity to be found, structured (partly monogamous)
merging wins. The design rule extends the one above: *merge widely when skills are additive; keep
structured sub-populations — island-style merging — when skills are rugged.* (Figure: `results/figS13_mating_breadth/E14.png`.)
*AI can do sex better than biology can.* Biology is stuck with two parents, mating roughly at random,
and cannot inspect an offspring before it is born. An AI has none of those limits. It can recombine
**many** parents at once; it can **choose** which parents to combine, for complementarity; and it can
**generate many candidate offspring and keep only the fittest**, screening them against reality before
committing. We call this **directed sex**, and in our simulations it converts the outbreeding-depression
catastrophe into a reliable gain: where blind recombination collapses on entangled problems, directed
recombination matches or beats the best parent every time. The language-model prototype shows the same
sign where it can: breeding many recombined Qwen offspring and keeping the one the verifier scores
highest beats the single averaged soup on hard tasks (and, unsurprisingly, does nothing extra on easy
tasks the soup already solves). This is a genuine advantage of engineered reproduction over the
biological kind, and we think it is one of the more useful ideas in the paper.
So the picture of §5 is: single-teacher copying is asexual and collapses (Muller's ratchet = model
collapse); the cure is to *ground* every birth in reality and to reproduce *sexually*, recombining
many complementary parents; and because AI sex can be many-parent, mate-chosen, and offspring-screened,
it is not merely a hedge against collapse but an engine that produces children fitter than any parent.
### The limit of sex: model speciation
Sex has a limit, and it is the sharpest new prediction this frame makes. Recombination works because
the parents are variations on a shared background; push two lineages far enough apart and their
combination is no longer viable. In biology this is **speciation** — the onset of **reproductive
isolation** — and its genetic mechanism is the **BatesonDobzhanskyMuller incompatibility** (BDMI):
an allele that arose in one lineage and an allele that arose in the other are each harmless on their
own background, but their *combination*, never tested by selection in either parent, is deleterious in
the hybrid (Dobzhansky, 1937; Muller, 1942; Orr, 1995). A merged model is precisely such a hybrid — a
single *recombinant* genotype, an F2-like object exposed to **recombination load**, not a hybrid-vigour
F1 — so the theory predicts a specific trajectory as two models diverge: **compatible → outbreeding
depression → hybrid inviability**.
We built this as an explicit model (a companion result). Two lineages descend from a common ancestor,
each substituting a *disjoint* set of loci — so each parent is adapted and neither carries an
incompatibility — and a fraction of the cross-lineage locus pairs are BDMIs that fire only when a hybrid
inherits *both* derived alleles. Sweeping the divergence between the parents reproduces the predicted
curve exactly: hybrid fitness tracks the parents while they are compatible, then peels off, peaks, and
crashes below the ancestor (an inviable hybrid). Three things fall out, and they are the contribution:
1. **The isolation cliff, and what moves it.** The divergence at which merging fails is not fixed: it
arrives *earlier the more epistatic the capability landscape*. In the model the reproductive-isolation
rate at high divergence rises from ~0 to ~0.5 as the density of incompatibilities grows. This is the
paper's distinct, falsifiable claim — **at matched divergence, mergeability is governed by epistasis,
not by divergence alone** — and it is exactly the axis that the machine-learning predictors of merge
success (which are all divergence/geometry measures) do not have.
2. **The snowball.** The number of incompatibilities grows with the *square* of the divergence
(Orr & Turelli, 2001), so hybrid fitness falls *super-linearly*: divergence is punished faster than
it accrues. Merge compatibility does not decay gently; it falls off a cliff.
3. **The design rule.** *Before merging, weigh divergence against the ruggedness of the shared
capability landscape; past the cliff, do not merge — route* (the engineering echo of allopatry:
keep the specialists reproductively separate and select among them instead of hybridising).
This is where a geneticist's lens earns its keep. The machine-learning literature has *observed* that
increasing specialisation eventually breaks merging and that one should then route rather than fuse
(Pari et al., 2024; Zhou et al., 2026), and part of the apparent incompatibility between independently
trained models is a coordinate artefact removable by aligning neurons (Git Re-Basin — Ainsworth et al.,
2022). What the frame adds is the *theory* of the phenomenon they observe: its functional form, its
super-linear (snowball) onset, and its dependence on epistasis — merge failure as a DobzhanskyMuller
event. (Figure: `results/fig5_speciation_bdm/E12.png`.)
**The real-weight confirmation.** The obvious objection to the analytic model is that its
"incompatibility" is a re-labelled loss barrier, and loss barriers between independently trained
networks are famously a *coordinate* artefact — two nets that learned the same function in a permuted
basis look incompatible until their neurons are aligned (Git Re-Basin), and recent work shows that
symmetry groups *richer* than permutations remove still more of the barrier (functionality-preserving
rescalings and rotations — Scaling LMC, 2026; neuron-identifiability approaches). We therefore ran the
experiment the objection demands, in real trained weights, aligning modulo the **full**
function-preserving unit symmetry group of the architecture (per-unit positive rescaling composed with
permutation — for a plain ReLU network, all of it). Two small MLPs are forked from a shared MNIST base,
trained, weight-averaged, and their linear-mode-connectivity error barrier is measured *before and
after* alignment; the after-alignment **residual** is the part of the incompatibility that no
re-coordination can explain away. The decomposition is clean (Figure:
`results/speciation_real/speciation_real.png`): two nets trained *from different random initialisations
on the same task* have a real naive barrier that alignment removes almost entirely (residual ≈ 0.001,
and the aligned merge performs at parent level) — same species, different basis, the canonical Re-Basin
result, which also proves the aligner works. Two nets that learned *conflicting* label maps have a
large barrier of which the full symmetry group removes **essentially nothing** (0.502 → 0.497) —
a conflict-associated barrier the tested alignment leaves largely unchanged — supporting a
functional-conflict interpretation without proving optimal alignment (control recovery validates a
special case; the removable share is a lower bound, the residual an upper bound). It also carries a floor no future alignment
method can breach: models loyal to label maps that conflict on a fraction *μ* of inputs cannot both be
served by *any* single merged model, which must err at rate ≥ *μ*/2 against at least one parent
(SI proposition). Sweeping the fraction of conflicting classes traces the **isolation cliff in real
weights**, now readable directly as *hybrid fitness*: the residual barrier climbs monotonically while
the merged model's accuracy falls from 0.97 to 0.03 — E12's compatible → depression → inviability
trajectory, measured.
**And its honest converse: speciation must be provoked; it did not emerge.** A true
DobzhanskyMuller incompatibility is *emergent* — each lineage's changes harmless alone, incompatible
only in combination — whereas the conflict condition above *imposes* contradiction. So we pre-registered
the emergent test: fork two children from a shared base and let them diverge with **no conflicting
training signal anywhere** — one pair as complementary class specialists (one child trains only on
digits 04, the other only on 59), one pair with divergent input conventions (views shifted in
opposite directions) — out to divergences 6.4× the base training. The result is the second
pre-registered reading, and it sharpens the theory's scope rather than confirming its most dramatic
form: the residual barrier is **0.000 at every divergence in both conditions**, and far from failing,
the merge *rescues* the two specialists — each parent decays toward ~0.50 on the full task
(catastrophically forgetting the classes it no longer sees) while the merged model holds ~0.95
throughout, a sustained FisherMuller rescue at zero barrier. In real weights, at least in this regime
of shared ancestry and compatible tasks, **reproductive isolation requires functional conflict; it does
not arise spontaneously from divergent specialisation.** The design rule sharpens accordingly: *merge
freely across divergently-specialised lineages of shared ancestry — what speciates model populations is
conflicting conventions, not specialisation per se.* Whether long-horizon over-specialisation erodes
mergeability at language-model scale — as the empirical merging literature hints (experts trained
longer merge worse under averaging) — is exactly the next tier's question, and the theory now makes the
prediction crisp: it should depend on whether extended training induces *conflicting conventions on
shared circuitry*, not on divergence time itself.
**What these experiments do and do not establish.** Stated at exactly the strength of the evidence:
they establish that *some merge failures reflect incompatible functional requirements rather than a
mismatch of coordinates* — a residual that survives the full unit-symmetry group of the architecture
tested, rises with functional conflict, and is absent under compatible specialisation. Three
qualifiers. First, the impossibility at the heart of the conflict condition — one deterministic model
cannot satisfy two contradictory answer conventions — is information-theoretic and needs no population
genetics; what the genetic frame adds is *structure around it*: which divergences generate such
conflicts, the prediction that epistasis rather than distance sets the cliff's position, and the
snowball's super-linear onset — the latter two verified so far only in the analytic model, and
therefore carried as **hypotheses at the neural tier, not results**. (On the snowball, one more
distinction: super-linear growth in the *number* of incompatibilities does not by itself entail a
sharp *performance* cliff — that needs the link from incompatibility count through effect sizes to
measured performance, which the analytic model supplies under its assumptions and any neural test
must establish separately.) Second, our alignment removes the symmetries we enumerate for this
architecture class, and exactly recovering a permuted-and-rescaled copy validates a special case
rather than proving global optimality for independently trained networks — so the removable share is
a lower bound and the residual an upper bound; richer transformation families for other architectures
could reapportion the split, though not below the conflict floor. Third,
"unmergeable" here means by aligned linear interpolation of weights — a barrier to that operator does
not preclude every conceivable recombination method (routing, for one, sidesteps it by not blending).
Emergent DobzhanskyMuller incompatibilities in real weights remain the flagship *hypothesis* of this
programme: our tested regimes found none, which bounds where they can live — longer horizons, shifted
data distributions, capacity pressure — and the decisive experiment (predicting merge success *before*
merging from an operational epistasis measure, against geometry- and gradient-based predictors) is
posed in the closing section.
One question remains, and the rest of the paper is largely about it: recombination combines what the
parents kept — but *who decides what each parent keeps, and which offspring are worth keeping?*
## 6. The second inheritance: letting "what is worth keeping" evolve
There are two answers, and the first is wrong. We could try to *design* the rule for what knowledge to
keep and pass on. But nobody knows that rule. "Keep the general, drop the particular" is a slogan, not
an algorithm: ask *which* generalisations, in *which* domain, at *which* grain, and the hand-written
rule falls apart. This is the deepest hole in the scheme, and it cannot be filled by decree.
The second answer is the one nature used: **do not design the selector — evolve it.** Let different
models carry different *policies* for what is worth keeping and combining. Let the policies that
produce more capable offspring spread; let the policies that produce weak offspring die out with their
lineages. The lineage's *taste* — its sense of what matters — is discovered by selection, not imposed.
So **two things are inherited, on two channels.** The *content* passes down directly: an offspring
receives its parents' knowledge (this is the "Lamarckian" channel — the inheritance of things acquired
during a lifetime, which biology forbids for genes but culture allows for ideas). The *selection
policy* — what to keep, whom to breed with, which offspring to screen for — is itself inherited, varies
between models, and survives in proportion to the success it produces. That second channel is
**Darwinian**. The architecture is therefore both at once: Lamarckian in *what* it transmits, Darwinian
in *what it keeps*. Evolutionary theorists call this structure *dual inheritance* and identify it as
the engine of human culture (Boyd & Richerson, 1985); philosophers of science describe scientific
knowledge itself as growing this way, by conjecture and **refutation** (Popper, 1959; Campbell, 1974;
Hull, 1988).
The closure that makes this fit together, rather than merely sound nice: Darwinian selection needs a
*selection pressure* — something that decides which policies win. That pressure is already in the
design. What tells a lineage its taste was good? The success of its offspring **against reality**. The
reality-check that stops collapse (grounding, §5) and the fitness signal that drives the evolving taste
turn out to be the *same thing*, seen from two sides.
## 7. The central danger: fitness is not truth
Introducing selection introduces selection's classic hazard, and it is severe enough to sink the whole
scheme if ignored. Evolution optimises, without mercy or foresight, for exactly what you *measure*
never for what you *meant*. (Economists and ML engineers know this as **Goodhart's law** and
*specification gaming*.) Get the fitness measure slightly wrong and the lineage will exploit the gap
with more ingenuity than any designed rule.
For a *knowledge* lineage there is a specific and nasty version. For ideas, the natural measure of
"fitness" is **how well they spread**, and a false-but-persuasive idea spreads beautifully. Human
intellectual culture is full of highly transmissible falsehoods; confident nonsense out-competes hedged
accuracy in almost every human forum. Turn Darwinian selection loose on models without care and it will
breed a lineage optimised for *persuasiveness* — fluent, compelling, and wrong. That is model collapse
with an optimiser behind it, actively seeking the cliff.
Only one thing makes fitness track truth rather than appeal: **being judged against a reality that can
say no.** Fitness must be predictive success under *intervention* — did the model's knowledge correctly
anticipate what the world would do when acted upon — and not approval, fluency, or a benchmark score,
each of which can be gamed. This is why the reality-check is load-bearing twice over: it is both the
anchor that stops passive collapse *and* the only thing that keeps the evolving taste honest.
The second danger is **convergence**, and beating it takes work at two separate levels, because
selection can only preserve variety that already exists — the variety must first be *supplied* and then
*kept*.
- **Supply.** A lineage that learns only from an accredited elite has a monoculture for a source: the
"best" experts are, almost by definition, the ones who won the consensus, so the incoming variation
is narrow from the start. The society must therefore learn, deliberately and from the beginning, from
the **outliers and the heterodox** as well as the credentialed — not out of fairness, but because in
evolutionary terms diverse founders are the raw material without which nothing downstream can adapt.
- **Preserve.** Even given varied input, plain fitness-*maximising* selection converges — it drives
every lineage toward the single current best and fixes it, extinguishing the rare specialists. The
fix is well established: **quality-diversity** selection, which rewards being *good* and being
*different* at once (novelty search and MAP-Elites — Lehman & Stanley, 2011; Mouret & Clune, 2015),
keeping complementary specialists alive rather than collapsing onto the champion. In our simulations
this is decisive: greedy "keep-the-best" selection collapses a population's diversity almost at once
and gets stuck at a mediocre answer, while quality-diversity selection keeps the specialists that
sexual recombination then needs as parents.
The two levels meet at reproduction. Multi-parent recombination (§5) is the *vehicle* by which the
diversity this selection preserves actually enters the next generation: an offspring drawn from
complementary parents inherits the standing variation the selector kept alive, recombined into one new
model. Supply the variety from the human side; preserve it on the selection side; recombine it into
each generation on the reproduction side. Remove any of the three and the lineage converges on its own
first guess.
## 8. A society needs institutions, not just specialists
One requirement is easy to overlook and fatal to omit. The easy part of a society is specialisation.
The *hard* part — which human civilisation took millennia to build — is the set of **institutions that
let fallible specialists combine without each re-verifying everything**: reputation, replication,
credentials, and above all **peer review**. These are error-correction protocols, and they exist
because a group of unreliable specialists left to reinforce one another is *more* wrong than any member
alone.
This is precisely where current multi-agent AI fails: set several models to confer and they tend to
agree sycophantically and confabulate in committee, because they have all the specialisation and none
of the institutions. A multigenerational society must specify not only how models learn, reproduce, and
are selected, but how they *check* one another — how a claim is challenged and a mistaken model loses
standing *before* its error is recombined into offspring and inherited. Peer review is itself a
reality-check of the kind §7 demands — an institutional stand-in for reality's "no," to be used where
direct intervention is slow or costly.
## 9. The lineage must stay open to reality
A society of models, however many generations deep, shares one hard limit: it has only ever *read*.
Its whole inheritance is a record of things that were said. In the vocabulary of causal reasoning
(Pearl, 2009), it lives on the bottom rung of the **ladder of causation** — observation — and no amount
of observation reaches *intervention*. Watching underdetermines doing; correlation does not contain
causation, at any scale.
Only intervention — reaching out and changing the world to see what happens — climbs the ladder, and a
language model cannot intervene. This is what humans and their instruments supply, and the contribution
is not "truth" but **constraint**: reality's unique gift is that it can say **no**. Text offers only
more opinion; an experiment delivers a refusal no consensus can overturn. As §§67 argued, that refusal
does double duty — it is both the anchor that prevents collapse and the fitness signal that lets the
lineage's evolving taste select for truth rather than persuasion.
Two honest riders. First, the human reality-signal is *dirty*: people supply results warped by
publication bias, incentive, and occasional fraud — which is exactly why the error-correcting
institutions of §8 must sit at the humanmachine boundary, screening the signal before it selects.
Second, humans are the *current* supplier of intervention, but the actuator half is being automated
(autonomous laboratories already close the designbuildtest loop). What looks durable in the human
role is therefore not the hands but the **choice of what to test and which refusals matter** — the
part of the fitness function that encodes *what is worth persisting*, as opposed to what merely *can*
persist. We flag, without resolving, that a partnership stays mutual only while both sides supply
something the other cannot.
## 10. Why it is cheap
A practical fact turns this from thought experiment into buildable proposal: **the architecture almost
never re-pays for the one genuinely expensive thing in AI — pre-training.** (The single exception,
periodically re-minting the base, is §11, and it is rare enough to be an amortised footnote.)
Training a foundation model from scratch consumes trillions of words and a fortune in compute. This
design does none of that per generation. Every model is *born* from an existing open-weight model that
already paid that cost; specialising one is a small patch trained in hours on a single consumer GPU;
running the society is ordinary inference; and reproducing — recombining parents into a child — is, in
the model-merging case, cheaper still, because it can be done directly on the weights with no retraining
at all (Akiba et al., 2024). Selection does cost more — you must run *populations* and discard the
unfit — but that is a multiplier over an already-cheap unit, not over a foundation-model budget.
The economics work only with **open-weight** models, for reasons practical and legal at once: you must
be free to inspect, modify, and redistribute the weights, and most proprietary licences forbid using a
model's outputs to train another — which is exactly what reproduction here does. This is not ideology
bolted on; it is a structural constraint, and a democratising one, since it puts the whole architecture
within reach of a single laboratory.
## 11. Can it grow forever? Consolidating knowledge back into the base
One question the design has assumed away: can the lineage accumulate *without end*? The individual is
bounded, and that is the clock. But the lineage seemed unbounded — each generation simply starts a
little ahead. Look closer and a second budget also fills.
Every new model is a pristine base plus an inherited **soft** delta — the acquired knowledge carried in
added patches rather than baked into the frozen core (§3). That soft delta is what makes the lineage
multigenerational; it is also what cannot grow forever cheaply. Stacked patches are not free: they slow
inference, and past some depth the accumulated delta is better *consolidated* than carried. The lineage,
too, matures.
The fix is the same operation, one level up. When a lineage's acquired knowledge has proven stable
across enough generations, **re-mint the base**: distil the accumulated soft inheritance into the
*weights* of a fresh foundation-scale model — a new base born already *natively knowing* what took many
generations to acquire in patches. The soft budget resets; the next epoch begins from a richer floor.
What was hard-won and *learned* becomes cheap and *innate*.
The pattern **echoes the Baldwin effect** (Baldwin, 1896; its clean computational demonstration is
Hinton & Nowlan, 1987): knowledge acquired and re-learned every generation eventually becoming part of
the innate endowment. We use the echo advisedly — Baldwin's mechanism is *selection* favouring
genotypes that learn the trait ever more easily, whereas re-minting is direct distillation, a
deliberate engineering shortcut through the same soft-to-innate valve. The valve is the point: two
substrates, the soft learned patches and the hard base weights every model is born with, with a
controlled passage between them.
Three honest riders, because re-minting is the most consequential step in the scheme:
- **Cost.** This is the one step that re-pays part of the pre-training bill, breaking §10's cheapness
*locally*. It is bearable only because it is *rare*, amortised over many cheap generations, and is
continued training from the lineage's own rich outputs rather than a de-novo run.
- **Irreversibility (of the lineage, not the archive).** A digital system can, of course, keep every
old base on disk — nothing forces deletion, and archives should be kept. The irreversibility is
*operational*: once the lineage's production base, training mixtures, and selection all run downstream
of the re-minted weights, a quiet collapse baked into them propagates to every descendant, and the
archived ancestor helps only if some process still compares against it — which nothing in the loop
does by default. In our minimal models a collapsed-then-re-minted lineage locks in its loss exactly
this way, and a cheap safeguard prevents it: **re-mint only while the lineage is demonstrably diverse
and healthy** (and keep an audit that diffs against the archived ancestor), never as a rescue for a
line already drifting. It is the sharpest instance of the human seat of §9 — choosing what no future
generation will think to question.
- **Speciation.** A re-minting is a founder event. Different laboratories, re-basing on different
criteria, will mint divergent bases; the lineage branches. This is not a defect but *adaptive
radiation*, and it is exactly what open weights make possible. The society grows not as one heavy
trunk but as a branching tree of bases.
So the honest answer to "can it grow forever?" is: **the architecture removes the *storage* obstacle
to indefinite accumulation** — nothing is retained without bound anywhere, and consolidation resets
the soft budget each epoch — but that is a statement about bookkeeping, not a demonstration of
unbounded capability growth, which no fixed-capacity system can promise and our finite models
(deliberately scoped as "effectively open-ended relative to the sample size, not astronomically
open-ended") do not test. What the design claims is the weaker, defensible thing: at no level does a
full store force the lineage to stop learning.
## 12. One process, four timescales
Step back and the parts resolve into a single idea running at four nested speeds. The **vertical**
motion is transmission — the selective passing-down of hard-won knowledge:
1. **Within one model, over a working life:** experience is consolidated from fast, episodic memory
into slow, durable weights, without catastrophic loss.
2. **Between generations, at maturity:** mature models reproduce — recombined into a fresh one.
3. **Across many generations:** each generation inherits the compressed achievements of the last and
builds on them.
4. **Across epochs:** a proven lineage's accumulated soft inheritance is consolidated into the weights
of a re-minted base, becoming innate.
The first and last are the *same operation at opposite ends of the scale* — a fast/soft store
consolidating into a slow/hard one — one running overnight inside a single model, the other across an
epoch inside a whole society. The **horizontal** motion is selection — Darwinian selection acting across
the population at each timescale, on the policies that govern what gets transmitted, with reality as the
fitness function and diversity-preservation keeping the specialists alive.
The same three rules govern all of it: **reproduce by recombining, not by copying, or you decay;
preserve the disagreements and the surprises, or you converge; and anchor fitness to a reality that can
refute, or you evolve toward what is merely convincing.**
## 13. What we built, what we found, and what is still open
The previous drafts of this paper promised a "companion paper" that *would* make this concrete. That
work now exists — mostly as a set of **minimal, laptop-reproducible models**, with a first bridge to
**real language models** (a LoRA-merge prototype, up to 7B on a GPU cluster) — and it is worth stating
plainly what it does and does not show. (A separate results document gives the numbers; here is the
shape.)
**What we built and found.**
- *An exact account of collapse.* Because generational training is the WrightFisher drift process, we
can check a simulator against century-old closed-form formulas, and it matches them to a fraction of
a percent. Collapse is not argued by analogy; it is derived.
- *The cheap-grounding result, and its limit.* A few percent of verified real data holds on to most of
a lineage's diversity indefinitely — but not the deepest tail, which needs recombination. This is
what makes a continually-learning society economically plausible rather than a data-hungry fantasy.
- *"Merge, don't average."* Combining several teachers by *averaging* their outputs — the obvious thing,
and what a "model soup" does — mathematically cancels the benefit of having several teachers. A
*merge* that keeps each item's strongest source realises it. Most current multi-model setups get this
wrong by default.
- *Collapse and its cure in real trained networks, and on real images.* We reproduced the same effects
in small recurrent and feed-forward networks and in a generator of handwritten digits (MNIST), where
a model trained on its own output collapses to a single blurred digit while a little grounding keeps
all the styles alive. An honest wrinkle we had to report: real neural networks *smooth*, so the naive
diversity metric misleads, and the right measure is distance-from-truth.
- *Sex that beats the parents, and when it doesn't.* In evolutionary simulations, recombining
complementary specialist models produces a model fitter than any parent (the FisherMuller effect),
climbing toward the best-possible combination as more, more-diverse parents are added — while
averaging and best-single-parent plateau below. On *entangled* problems, blind recombination instead
produces below-parent offspring (outbreeding depression) — and *directed* recombination (choose mates,
screen offspring, unbounded parents) reliably fixes it. This is the concrete evidence for the paper's
central reframing.
- *The mating system, not just the mating.* Sweeping how *widely* models recombine — from monogamous
(local, structured) to promiscuous (panmictic) — against landscape ruggedness, the best breadth
**shrinks as skills get more entangled**: wide, promiscuous merging wins on additive landscapes, but on
rugged ones it prematurely converges to a worse champion and an intermediate, structured breadth wins,
because promiscuity monotonically destroys the diversity a rugged search needs. A merging-native design
axis — *merge widely for additive skills, keep island-structured sub-populations for entangled ones*
that the model-merging literature, which assumes panmixia, does not have.
- *The recombination claims, in real language models — with a sharp condition.* Merging LoRA-specialised
Qwen models (up to 7B on a GPU cluster) produces a generalist that beats every specialist parent
(FisherMuller, for real); and keeping parents intact and *routing*, or *breeding and screening*
offspring, beats the naive average — but *only when the task leaves headroom*. On easy tasks a strong
model's plain average is already at the ceiling and the refinements add nothing; on hard tasks the
average dilutes a specialist below even the best single parent, and the union-preserving operators win
clearly. The practical rule, stated qualitatively: these tricks pay off where the naive average falls
short of attainable performance, and add nothing where it does not (a quantitative form is untested).
This is a prototype (three task families, one seed), so we read it as
signs, not magnitudes; the *whole grounded society* on a language model remains the open step.
- *The whole society, and why every part is needed.* In a population evolving on a "reality" landscape,
the full system — grounding + sexual recombination + preserved diversity — climbs to the top while
keeping its specialists. Remove *grounding* and it collapses into a confident, wrong consensus (a
direct analogue of training on the internet's growing crowd of AI-generated text); remove *sex* and it
gets stuck; remove *diversity* and it converges too fast to a worse answer. Each removal fails
differently; only the whole system climbs. This is the closest thing we have to a test of the actual
thesis, rather than of the borrowed scaffolding around it.
### The claims at a glance: status, assumptions, evidence, limits
Because a perspective of this breadth risks blurring what is proved, what is measured, and what is
proposed, here is the ledger of the load-bearing claims — each labelled **exact** (closed-form in the
minimal model), **empirical** (measured in trained systems), or **hypothesis** (stated with a
falsifier, not yet established):
| Claim | Status | Key assumptions | Evidence | Known limits |
|---|---|---|---|---|
| Collapse = WrightFisher drift (minimal model) | Exact (diagnosis conceded to prior work) | Knowledge = categorical distribution; refit = resample | Closed forms reproduced to <0.5% | Real learners add a signed, architecture-specific estimator bias (measured) |
| Grounding = immigration; critical real-data fraction ≪ 1 | Exact + empirical sign | Fresh samples from a fixed, non-drifting truth | Exact `H_eq`; `g*≈0.048`; sign holds in RNN/MLP/VAE and on MNIST | Deepest tail unrescuable at feasible budgets (`m 1/p`); sharp threshold softens in trained nets |
| "Merge, don't average" conservation | Exact **for the output-mean operator** | Rare-item regime; an oracle/verifier identifies the strongest source | E4 closed form + simulation; neural reproduction | Weight-averaging and routing are empirical cousins, not instances; budgets differ; bridge = the headroom rule |
| Offspring exceed every parent (FisherMuller) | Interpretation + empirical | Complementary (decorrelated) parents; verifiable fitness | E8 analytic; 7B LoRA merge beats every specialist on every family | LLM tier: 3 lexically-distinct families; multi-seed replication in progress |
| Outbreeding depression on rugged landscapes; operator design rule | Exact-model result; hypothesis at LLM scale | NK epistasis stands in for skill entanglement | E9E10; directed selection rescues | Not yet mapped onto a real task-entanglement measure |
| Optimal mate-pool breadth shrinks with ruggedness | Exact-model result; hypothesis for merging populations | Ring population, local selection | E14 | Phenomenon known to island-model evolutionary computation; our contribution is the mapping and the diversity/mean decomposition |
| Merge failure decomposes into coordinate artefact + functional residual | Empirical (MLP tier; LLM tier in progress) | Alignment enumerates the architecture's unit symmetries | Full-symmetry residual ≈ 0 (compatible) vs ≈ naive (conflict); cliff in hybrid fitness | Scoped to aligned linear interpolation; conflict floor is information-theoretic, not genetic |
| Epistasis (not divergence) sets the cliff; snowball onset | Exact-model result; **hypothesis** at the neural tier | BDM incompatibility structure | E12 | Snowball count ≠ performance cliff without the effect-size link; neural test outstanding |
| Pre-merge functional disagreement predicts merge penalty | Empirical, within a controlled grid (0.5B, 13 conditions × 3 seeds) | Constructed conflict/overlap/duration axes; oracle-potential outcome (pre-registered; ordering sensitive to reference) | Clustered CIs exclude 0; held-out LOCO ρ≈0.4; selected geometry baselines ≈ 0 | Head-to-head predictor differences not individually significant; only selected baselines; generalisation to real task pairs open |
| Confidence weighting improves rank prediction over raw disagreement | **Not supported** (pre-registered internal prediction) | — | Paired Δ\|ρ\| ≈ 0.02, CI [0.13, +0.06] | Weighting does double the conflict-vs-compat level contrast |
| The predictor improves budget-matched operator choice | **Open** | — | Soup-vs-route gap readout noise-dominated at 0.5B | The practical payoff; untested |
| Emergent speciation without conflict | **Not observed** (pre-registered) | Shared ancestry, compatible tasks, tested divergences | E13b: residual 0.000; merge rescues specialists | Bounds the hypothesis; longer horizons/distribution shift/capacity pressure untested |
| Grounding + sex + diversity complementary (each ablation fails distinctly) | Analytic-model result; hypothesis at LLM scale | Conformity stands in for self-consumption; general joint necessity not established | E11 four-arm ablation | The full grounded LLM society is unbuilt; alternative schemes untested |
**What is borrowed, and what is ours.** We are deliberate about the ledger, because the surrounding
literature is crowded and a reader deserves to know exactly where the line falls. **Conceded as prior
art:** (a) *model collapse is genetic drift* — derived independently and cleanly (Riis, 2026; the
WrightFisher collapse literature following Shumailov et al., 2024; the closed-form first-extinction
law of Benati et al., 2025; the quantitative-trait account of Yoon et al., 2025); (b) the empirical
facts that a merged model can *beat its parents*, that *decorrelated* parents merge better, and that
*naive averaging is inferior* to sign-reconciled or routed merges (model soups, TIES, DARE,
mixture-of-experts routing); (c) that a *population* of merging or self-improving models can climb
(GENOME, M2N2, Multiagent Finetuning, the DarwinGödel Machine); (d) that merge success has
machine-learning-native *predictors* — interpretable pairwise metrics (Zhou et al., 2026),
capacity/rate-distortion accounts of merging collapse (Cao et al., 2026), and stability/scaling
analyses of multi-task degradation; and (e) that verifier-screened synthetic data can avert collapse
(Yi et al., 2025) — the statistical cousin of our grounding operator. We claim none of these.
**Ours** is the framework those results invite: a **population-genetics of sex** applied to model
societies, generative where the incumbents are empirical. Concretely — the **"merge, don't
average" conservation law** (recombination preserves the union; blending inheritance cancels it),
derived not observed; **FisherMuller** named and used to explain *why* offspring exceed parents;
**outbreeding depression on rugged/epistatic landscapes**, which turns "when does merging help vs hurt"
from a thing you must run a search to discover into a thing the landscape's ruggedness *predicts*, with
the operator-choice design rule that follows (average / union-route / directed-select); **grounding as
migrationdrift balance**, giving a critical real-data fraction and a phase boundary a closed
self-consuming loop cannot have; **directed sex** as the distinctly-AI advantage (unbounded parents,
offspring preview, mate choice); and the **integrated society** whose operators make
*complementary, distinctly-failing contributions* in the tested model (general joint necessity is not
established). The value-add over the machine-learning-native merge theory is that ours predicts *which
operator to use and when it will backfire*, not merely how fast quality decays. And it opens — and
begins to occupy — a question nobody has framed: **model speciation**, the population-genetics of
*reproductive isolation* (BatesonDobzhanskyMuller incompatibilities) as the account of *when two
models are too diverged to be merged at all*. We model it explicitly (§5), predicting the
compatible → outbreeding-depression → inviability curve, its super-linear (snowball) onset, and its
control by epistasis rather than divergence alone — the one place the merge literature has phenomena
(Pari et al., 2024; Zhou et al., 2026) but no theory — and we confirm it in real trained weights, where
a merge barrier survives alignment under the *full* function-preserving symmetry group (not only
Re-Basin permutations) as a residual, functional reproductive isolation with an information-theoretic
floor — together with the pre-registered emergent converse: absent conflicting training signals,
divergently-specialised lineages of shared ancestry showed *no* isolation at any divergence tested, the
merge instead rescuing the forgetting specialists (isolation must be provoked; specialisation alone did
not speciate). In one sentence: the field agrees on the disease
and tinkers at the cure with evolutionary metaphors; we bring the evolutionary *theory*, and it makes
falsifiable predictions — a merge-compatibility cliff among them — that the metaphors do not.
**What is still open — honestly.** The old hole (what to select) we fill in kind: don't design the
selector, evolve it. But the hole has *moved*, not closed, and the new one is harder: **the fitness
function** — what reality-anchored measure selects for *truth* without also selecting for *persuasion*,
given that in our own species the two have been at war for the whole history of ideas. Alongside it:
the **institutions** that let contemporaries correct one another before error is inherited (§8), which
we do not solve; and the **calibration** of everything the results left as knobs — how many parents,
how complementary, at what ratio of inherited-to-real data, and how healthy a lineage must be before
its knowledge is safe to make irreversibly innate. These are, at least, *measurable* — which is the
difference between an open problem and a hole. And the largest gap of all: the *recombination* claims
now hold in real language models, but the *society* — the grounded, diversity-preserving, continually
reproducing loop — does not yet. The real test is to build that whole system out of actual open-weight
language models, and see whether all the signs survive contact with a system too big to write down.
The operators, checked; the living society, next.
---
## Selected references
- Akiba, T., Shing, M., Tang, Y., Sun, Q., & Ha, D. (2024). Evolutionary optimization of model merging recipes. *Nature Machine Intelligence.* (See also Sakana AI's M2N2, "Model Merging of Natural Niches.")
- Baldwin, J. M. (1896). A new factor in evolution. *The American Naturalist.*
- Boyd, R., & Richerson, P. J. (1985). *Culture and the Evolutionary Process.*
- Campbell, D. T. (1974). Evolutionary epistemology. In *The Philosophy of Karl Popper.*
- Fisher, R. A. (1930). *The Genetical Theory of Natural Selection.*
- French, R. M. (1999). Catastrophic forgetting in connectionist networks. *Trends in Cognitive Sciences.*
- Hinton, G. E., & Nowlan, S. J. (1987). How learning can guide evolution. *Complex Systems.*
- Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network. *arXiv:1503.02531.*
- Hu, E. J., et al. (2021). LoRA: low-rank adaptation of large language models. *arXiv:2106.09685.*
- Hull, D. L. (1988). *Science as a Process.*
- Kauffman, S. A., & Levin, S. (1987). Towards a general theory of adaptive walks on rugged landscapes. *Journal of Theoretical Biology.* (The NK model.)
- Lehman, J., & Stanley, K. O. (2011). Abandoning objectives: evolution through the search for novelty alone. *Evolutionary Computation.*
- Mallya, A., & Lazebnik, S. (2018). PackNet: adding multiple tasks to a single network by iterative pruning. *CVPR.*
- McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex. *Psychological Review.*
- McCloskey, M., & Cohen, N. J. (1989). Catastrophic interference in connectionist networks. *Psychology of Learning and Motivation.*
- Minsky, M. (1986). *The Society of Mind.*
- Mouret, J.-B., & Clune, J. (2015). Illuminating search spaces by mapping elites (MAP-Elites). *arXiv:1504.04909.*
- Muller, H. J. (1932). Some genetic aspects of sex. *The American Naturalist.* (The advantage of recombination.)
- Muller, H. J. (1964). The relation of recombination to mutational advance. *Mutation Research.* (Muller's ratchet.)
- Pearl, J. (2009). *Causality: Models, Reasoning, and Inference* (2nd ed.).
- Popper, K. (1959). *The Logic of Scientific Discovery.*
- Riis, S. (2026). Drift and selection in LLM text ecosystems. *arXiv:2604.08554.*
- Rusu, A. A., et al. (2016). Progressive neural networks. *arXiv:1606.04671.*
- Shumailov, I., et al. (2024). AI models collapse when trained on recursively generated data. *Nature.*
- Wortsman, M., et al. (2022). Model soups: averaging weights of multiple fine-tuned models. *arXiv:2203.05482.*
- Wright, S. (1931). Evolution in Mendelian populations. *Genetics.*
*The evolution of sex (the geneticist's canon this paper draws on):*
- Barton, N. H., & Charlesworth, B. (1998). Why sex and recombination? *Science.*
- Otto, S. P., & Lenormand, T. (2002). Resolving the paradox of sex and recombination. *Nature Reviews Genetics.*
- Kondrashov, A. S. (1993). Classification of hypotheses on the advantage of amphimixis. *Journal of Heredity.*
- Dobzhansky, T. (1936); Muller, H. J. (1942). BatesonDobzhanskyMuller incompatibilities (reproductive isolation).
- Livnat, A., & Papadimitriou, C. (2016). Sex as an algorithm: the theory of evolution under the lens of computation. *Communications of the ACM 59(11).* (The theory-of-computation precursor: recombination selects for mixability.)
*The 20252026 landscape this paper positions against:*
- Subramaniam, V., Du, Y., Tenenbaum, J. B., Torralba, A., Li, S., & Mordatch, I. (2025). Multiagent finetuning: self-improvement with diverse reasoning chains. *arXiv:2501.05707.*
- Zhang, J., Hu, S., Lu, C., Lange, R., & Clune, J. (2025). Darwin Gödel Machine: open-ended evolution of self-improving agents. *arXiv:2505.22954.*
- *Nature-inspired population-based evolution of large language models* (GENOME/GENOME+). (2025). *arXiv:2503.01155.*
- Sakana AI (2025). Competition and attraction improve model fusion (M2N2). *arXiv:2508.16204* (GECCO '25).
- Yadav, P., Tam, D., Choshen, L., Raffel, C., & Bansal, M. (2023). TIES-Merging: resolving interference when merging models. *NeurIPS / arXiv:2306.01708.*
- Yu, L., Yu, B., Yu, H., Huang, F., & Li, Y. (2023). Language models are super Mario: absorbing abilities from homologous models (DARE). *arXiv:2311.03099.*
- Gerstgrasser, M., et al. (2024). Is model collapse inevitable? Breaking the curse of recursion by accumulating real and synthetic data. *arXiv:2404.01413.*
- Guo, D., Wu, J., & Yiu, S. M. (2026). Model collapse as cultural evolution. *arXiv:2605.23054.*
- Benati, M., Londei, A., Lanzieri, D., & Loreto, V. (2025). First-extinction law for resampling processes. *arXiv:2509.20101.* (Collapse onset = the WrightFisher first-extinction time.)
- Yoon, Y., Hu, D., Weissburg, I., Qin, Y., & Jeong, H. (2025). Model collapse in the self-consuming chain of diffusion finetuning: a novel perspective from quantitative trait modeling. *ICLR 2025 / arXiv:2407.17493.*
- Yi, B., Liu, Q., Cheng, Y., & Xu, H. (2025). Escaping model collapse via synthetic data verification. *arXiv:2510.16657.*
- Ainsworth, S., Hayase, J., & Srinivasa, S. (2022). Git Re-Basin: merging models modulo permutation symmetries. *arXiv:2209.04836.*
- Li, T., & Shen, Z. (2026). Scaling linear mode connectivity and merging to billion-parameter pretrained transformers. *arXiv:2606.23607.* (Symmetry groups richer than permutations remove more of the barrier.)
- Sharma, E., Roy, D. M., & Dziugaite, G. K. (2024). The non-local model merging problem: permutation symmetries and variance collapse. *arXiv:2410.12766.*
- Pari, J., Jelassi, S., & Agrawal, P. (2024). Collective model intelligence requires compatible specialization. *arXiv:2411.02207.*
- Zhou, L., Zhao, B., Yu, R., & Rodolà, E. (2026). Demystifying mergeability: interpretable properties to predict model merging success. *arXiv:2601.22285.*
- Cao, Y., Ran, D., Guo, Y., Wu, M., Chen, S., et al. (2026). An empirical study and theoretical explanation on task-level model-merging collapse. *arXiv:2603.09463.*
- Hu, Y., Yao, Y., Zhang, N., Chen, H., & Deng, S. (2024). Exploring model kinship for merging large language models. *arXiv:2410.12613.*
- Kozodoi, N., Afolabi, Z., & Butler, J. (2026). Are we merging the right models? Impact of expert training duration on model merging for LLMs. *arXiv:2607.11997.*
- Harris, K. D. (2026). A mathematical theory of evolution for self-designing AIs. *arXiv:2604.05142.*
- Chen, N., Tong, Y., Yang, Y., He, Y., Zhang, X., et al. (2026). Diversity collapse in multi-agent LLM systems: structural coupling and collective failure in open-ended idea generation. *arXiv:2604.18005.*
- Tanaka, H. (2026). When is collective intelligence a lottery? Multi-agent scaling laws for memetic drift in LLMs. *arXiv:2603.24676.*
*Still to engage in a full version: tacit knowledge (Polanyi) and human capital (Becker).*

View file

@ -1,570 +0,0 @@
# The Lamarckian Society
### AI that reproduces sexually: how a society of models can keep learning across generations instead of collapsing
*A perspective. Draft 5 — the companion to a set of minimal working models (now built).*
---
### A note on vocabulary (please read this first)
This paper sits at the meeting point of three fields, and it is written so that a reader from any
one of them can follow all of it. We therefore **spell out** each field's jargon the first time it
appears, even at the risk of belabouring the obvious for the specialist. A short glossary, in case
you skipped a definition:
- **Model collapse** *(machine learning)* — the degeneration that happens when you train a model on
data produced by earlier models, over and over: rare cases disappear and the model drifts toward a
bland average.
- **Distillation** *(machine learning)* — training a fresh "student" model on the outputs of one or
more "teacher" models, so the student ends up knowing a compressed version of what they knew.
- **Model merging** *(machine learning)* — combining several trained models directly, at the level
of their weights, into one — no retraining. (Think of it as breeding two models rather than
teaching a third.)
- **Genetic drift** *(population genetics)* — the random loss of rare variants that happens in any
finite population simply because not everyone leaves offspring. It is the neutral, no-selection
baseline of evolution.
- **WrightFisher process** *(population genetics)* — the standard mathematical model of drift. We
will claim, and show, that generational model-training *is* this process, not merely like it.
- **Recombination / sexual reproduction** *(biology)* — making an offspring by combining pieces from
more than one parent, rather than copying a single parent (which is *asexual* reproduction).
- **Muller's ratchet** *(population genetics)* — the way an asexual lineage, one that never
recombines, accumulates damage it can never undo. It is, we will argue, the same thing as model
collapse.
- **Catastrophic forgetting** *(machine learning / neuroscience)* — a neural network overwriting what
it knew when it learns something new.
We have tried to keep the big picture legible on every page, and to be candid about what is argument
and what is evidence. The evidence is mostly from **deliberately small models** — mathematics, small
neural networks, image generators, and evolutionary simulations. A first bridge to real language
models exists — a prototype that recombines LoRA-specialised Qwen models up to 7B on a GPU cluster,
which confirms the recombination signs (below) — but the *full grounded society* has not yet been
built on a large language model. We will say so repeatedly, because the gap matters.
---
## Abstract
We want AI that keeps learning across generations — the way a research field or a culture does,
each generation standing on the compressed knowledge of the last — rather than a single model trained
once and frozen. The obstacle is well known to machine-learning engineers as **model collapse**:
train each generation on the previous one's output and quality degrades, the rare cases vanishing
first. The central observation of this paper is that this failure is *reproduction gone wrong*, and
that biology already knows the fix.
Copying one model into the next — a "teacher" distilled into a "pupil" — is **asexual reproduction**.
Asexual lineages, in nature, decay: they accumulate errors they cannot undo (a process geneticists
call **Muller's ratchet**), and this decay is, mechanically, model collapse. The remedy nature found,
hundreds of millions of years ago, is **sex**: build each new individual by *recombining* several
parents, so it inherits a combination none of them had — and can be **fitter than any of its
parents**. We argue that a society of AI models should reproduce the same way: each new model
**recombined from many complementary "parent" models** (something the field already does, under the
name *model merging*), its selection anchored to **reality** (so it is judged against the world, not
against the consensus of other models), and its diversity actively preserved. With those three
ingredients — recombination, grounding in reality, and preserved diversity — a lineage does not merely
avoid collapse; it **climbs**, producing models better than any single ancestor while each specialty
is re-learned and surpassed.
AI has one advantage biology lacks: its "sex" has **no two-parent limit**, its mates can be **chosen**
for complementarity, and its offspring can be **screened before they are kept**. We call this
*directed sex*, and it turns recombination from a gamble into a reliable engine.
We support the argument with a set of **minimal models**: a mathematically exact account of collapse
and its cure; reproductions of the same effects in small trained neural networks and in a generator of
handwritten digits; and evolutionary simulations in which a population of models climbs a "fitness
landscape" that stands in for reality. In these, every claim above either holds or fails visibly, and
removing any single ingredient breaks the system in a distinct way. A first **language-model
prototype** then confirms the recombination claims in real weights — merging LoRA-specialised Qwen
models (up to 7B on a GPU cluster) produces a generalist that exceeds every parent, and shows the
sharp condition under which the "merge, don't average" refinement matters (below). The scope is
honest: these are existence proofs and design rules, and the eventual test is to build the *whole
grounded society* out of real language models. We also note where our diagnosis is no longer novel — the reading of
collapse as genetic drift has since been derived independently — and locate our contribution in the
**cure** rather than the diagnosis.
---
## 1. From a society in space to a society in time
The idea of many AI agents working together — a "society of mind" (Minsky, 1986), or today's
multi-agent systems — arranges intelligence across *space*: several specialists side by side,
dividing a task. This paper is about a different axis: *time*. Not a society that merely exists at
one moment, but one that **persists and renews across generations**, each new cohort of models
starting from the compressed knowledge of the last.
The unit that matters is therefore the **generation**, and the event that matters is **reproduction**:
the making of a new model from older ones. A single model, like a single mind, is bounded and
eventually stops improving. A *lineage* need not be. Human civilisation is not clever because any one
person is; it is clever because each generation inherits the distilled achievements of the previous
one and adds a little. We propose building AI the same way — and, crucially, getting the *reproduction*
right, because that is exactly where it can go wrong.
## 2. Why today's models cannot do this
Today's large language models have no life cycle. They are trained once, at enormous cost, then
**frozen** and deployed as a fixed artefact that does not learn from the people it serves. Learning
and doing are split into two eras with no bridge between them.
There is a real reason for the freeze. Updating a neural network on new information tends to overwrite
what it already knew — **catastrophic forgetting**, a problem understood since the late 1980s
(McCloskey & Cohen, 1989; French, 1999). Freezing avoids it by refusing to learn at all. The result
is a mind with no childhood, no growth, and no way to pass anything on. A lineage needs the opposite:
members that learn through their working lives, reach maturity, and hand on what they gained. So the
first requirement is a learner that can grow *safely*.
## 3. A learner that can grow without forgetting
The individual model needs two properties.
**It must not catastrophically forget.** Instead of overwriting its core as it learns, it keeps that
core frozen and only *readable*, and carves each new skill into freshly-added capacity beside it. In
machine learning this is called *parameter isolation* (progressive networks — Rusu et al., 2016;
prune-and-freeze — Mallya & Lazebnik, 2018; and, most practically, **LoRA** and other small trainable
"patches" bolted onto a frozen model — Hu et al., 2021). If the core is never altered, forgetting it
is not merely unlikely but structurally impossible. This is what lets a model accumulate a coherent
working life of expertise — the kind of stable knowledge worth passing on.
The brain offers a partial blueprint. *Complementary Learning Systems* theory (McClelland,
McNaughton & O'Reilly, 1995) — itself a response to the forgetting problem — describes two subsystems:
a **fast** store (the hippocampus) that grabs an experience in one shot, and a **slow** store (the
neocortex) that integrates regularities gradually without disruption. We do not lean on any particular
account of how the brain moves knowledge between them; the architecture needs only that *some*
periodic **offline consolidation** step exists, moving knowledge from the fast store to the slow one
when the system is idle. The machine version is clean regardless: the prompt is working memory, an
external database is the fast episodic store, the trained weights are the slow store, and consolidation
migrates the first into the last.
**It is bounded.** Because the model only ever *adds* capacity and freezes what it has, it eventually
fills up. In most designs that is a wall to dread. In ours it is a clock.
## 4. "Full" is maturity, not failure
Here is the pivot. A bounded learner that fills up has not broken. **It has grown up.**
Read the capacity limit as a life stage. A model is *born* as a freshly-schooled base — its general
education. It enters a **working life**, adding specialised knowledge as it does its job. And it
reaches **maturity**: the point where it has learned much of what one working life in its niche can
teach. Maturity is not the end of usefulness — it is the moment the model is most worth learning
*from*. So maturity is the cue to **reproduce**. The capacity ceiling that every other architecture
fights becomes, in ours, the metronome of the generations.
Everything now turns on how that reproduction is done — and this is where the paper's central claim
lives.
## 5. Reproduction: copying collapses, recombination climbs
Suppose a mature model simply teaches a fresh one — distillation, one teacher to one pupil, generation
after generation. This is the obvious design, and it fails, for a reason that is exactly the same in
machine learning and in biology.
**The machine-learning statement.** Training each generation on the previous generation's outputs is
the recipe for **model collapse**: the model forgets the improbable, loses the *tail* of the
distribution (the rare cases) first, and drifts toward its own most common output (Shumailov et al.,
2024). Worse for us, the very rule that makes distillation useful — *keep the general, drop the
idiosyncratic* — **is** tail-deletion by design. The operation that would power a cultural ratchet and
the operation that drives model collapse are the same act.
**The population-genetics statement (the same thing).** Represent a model's knowledge as a
distribution over discrete "items" — capabilities, facts, modes of behaviour. One generation is:
*draw a finite sample from the parent, and refit the child to it.* That finite-sampling step is
**mathematically identical** to **genetic drift** — the random loss of rare variants in a finite
population — described by the century-old **WrightFisher** model (Wright, 1931; Fisher, 1930). This is
not an analogy we find pretty; it is the same equations, and we use them as an exact check on our
simulations (the first of the minimal models below). Rare items go extinct first, roughly ten times
faster than common ones, precisely as drift predicts.
And single-teacher copying is **asexual reproduction** — cloning one parent. Nature already knows what
happens to an asexual lineage that never recombines: it accumulates damage it can never repair, a
one-way decline geneticists call **Muller's ratchet** (Muller, 1964). *Muller's ratchet is model
collapse.* Naming it that way is not decoration; it tells us where the cure is, because biology solved
this problem.
Two ingredients turn the collapse operation into a climb. Both are things nature does.
**First: do not reproduce "dry."** Model collapse is a property of a lineage fed *only* its own
output; the documented fix is that keeping some real data in the mixture arrests it (Shumailov et al.,
2024). We call that real data **grounding** — fresh contact with the world, verified against it. In
our minimal models, grounding is startlingly cheap: mixing in even a few percent of verified real data
holds on to most of the diversity indefinitely. But — an honest limit we found and did not expect —
grounding cannot save the *very rarest* items at any affordable budget; protecting an item of rarity
*p* needs a real-data budget that grows like 1/*p*. Grounding rescues diversity cheaply; it does not,
by itself, rescue the deep tail. Something else must. That something is sex.
**Second: reproduce sexually.** Instead of copying one parent, build each new model by **recombining
several** — a *sexual* rather than asexual birth. In machine learning this already has a name and a
working implementation: **model merging** (Akiba et al., 2024). Its importance here is not efficiency;
it is that recombination does something copying cannot. If several parent models have each specialised
on different parts of reality, each has kept alive rare knowledge the others lost. A recombined child
inherits the **union** of what its parents kept — not the tail-thinned *average* of a crowd of
near-identical copies. And here is the point that lifts sex from a safeguard to the engine of the whole
scheme, and the reason biology invented it:
> **An offspring recombined from complementary parents can be *fitter than any of its parents*.**
Geneticists call this the **FisherMuller effect** (Fisher, 1930; Muller, 1932): recombination brings
together, in one individual, beneficial variants that arose separately in different lineages, so the
child holds a combination none of the parents had. In our simulations this is exactly what we see —
recombining decorrelated specialist models yields a model that climbs toward the best-possible
combination, a genotype *no single parent possessed*, while the best single parent, and the naive
average of all of them (what the field calls a "model soup" — Wortsman et al., 2022), both plateau
well below. This is the concrete meaning of the paper's title claim, "the lineage climbs in general
knowledge; specialisation is re-earned each generation," and it is why the reframing from
teacher→pupil to *sexual reproduction* is not cosmetic: **copying can only recover a ceiling;
recombination can exceed it.**
This is no longer only a simulation. In a first language-model prototype — LoRA specialists on
disjoint task families, recombined and judged by an exact verifier — a merge of three specialist Qwen
models (7B, on a GPU cluster) **beats every single specialist**, overall and on every family: the
FisherMuller effect, in real weights. The same prototype pins down *when* the finer "inherit the
union, don't average" rule actually bites. Keeping each parent whole and **routing** each input to the
right one beats the tail-thinning average — but only when the task is hard enough to leave room to
lose: on easy tasks a strong model's plain average is already at the ceiling, so the crude soup is
fine, whereas on hard tasks the average dilutes a hard-won specialist so badly it falls below even the
best single parent, and routing wins by a wide margin. The rule is therefore precise: **the union
beats the average in exact proportion to how far the average is from the best attainable** — a caveat
that sharpens rather than weakens the claim, and that a practitioner needs before spending compute on
the fancier operator.
Two caveats keep this honest, and both are results, not hand-waving.
*Sex can backfire.* When the parents' skills are not cleanly separable but **entangled** — when the
value of one capability depends on which others are present (geneticists call this **epistasis**) —
blindly recombining two good models can produce a *worse* child, because recombination breaks up a
combination that only worked as a whole. Biologists call this **outbreeding depression**, and we
reproduce it: on "rugged" (highly entangled) problems, naive merging drops offspring below their
parents, and the more you mix the worse it gets. The design rule that falls out is simple: *merge
freely when skills are complementary; merge sparingly, and carefully, when they are entangled.*
*AI can do sex better than biology can.* Biology is stuck with two parents, mating roughly at random,
and cannot inspect an offspring before it is born. An AI has none of those limits. It can recombine
**many** parents at once; it can **choose** which parents to combine, for complementarity; and it can
**generate many candidate offspring and keep only the fittest**, screening them against reality before
committing. We call this **directed sex**, and in our simulations it converts the outbreeding-depression
catastrophe into a reliable gain: where blind recombination collapses on entangled problems, directed
recombination matches or beats the best parent every time. The language-model prototype shows the same
sign where it can: breeding many recombined Qwen offspring and keeping the one the verifier scores
highest beats the single averaged soup on hard tasks (and, unsurprisingly, does nothing extra on easy
tasks the soup already solves). This is a genuine advantage of engineered reproduction over the
biological kind, and we think it is one of the more useful ideas in the paper.
So the picture of §5 is: single-teacher copying is asexual and collapses (Muller's ratchet = model
collapse); the cure is to *ground* every birth in reality and to reproduce *sexually*, recombining
many complementary parents; and because AI sex can be many-parent, mate-chosen, and offspring-screened,
it is not merely a hedge against collapse but an engine that produces children fitter than any parent.
One question remains, and the rest of the paper is largely about it: recombination combines what the
parents kept — but *who decides what each parent keeps, and which offspring are worth keeping?*
## 6. The second inheritance: letting "what is worth keeping" evolve
There are two answers, and the first is wrong. We could try to *design* the rule for what knowledge to
keep and pass on. But nobody knows that rule. "Keep the general, drop the particular" is a slogan, not
an algorithm: ask *which* generalisations, in *which* domain, at *which* grain, and the hand-written
rule falls apart. This is the deepest hole in the scheme, and it cannot be filled by decree.
The second answer is the one nature used: **do not design the selector — evolve it.** Let different
models carry different *policies* for what is worth keeping and combining. Let the policies that
produce more capable offspring spread; let the policies that produce weak offspring die out with their
lineages. The lineage's *taste* — its sense of what matters — is discovered by selection, not imposed.
So **two things are inherited, on two channels.** The *content* passes down directly: an offspring
receives its parents' knowledge (this is the "Lamarckian" channel — the inheritance of things acquired
during a lifetime, which biology forbids for genes but culture allows for ideas). The *selection
policy* — what to keep, whom to breed with, which offspring to screen for — is itself inherited, varies
between models, and survives in proportion to the success it produces. That second channel is
**Darwinian**. The architecture is therefore both at once: Lamarckian in *what* it transmits, Darwinian
in *what it keeps*. Evolutionary theorists call this structure *dual inheritance* and identify it as
the engine of human culture (Boyd & Richerson, 1985); philosophers of science describe scientific
knowledge itself as growing this way, by conjecture and **refutation** (Popper, 1959; Campbell, 1974;
Hull, 1988).
The closure that makes this fit together, rather than merely sound nice: Darwinian selection needs a
*selection pressure* — something that decides which policies win. That pressure is already in the
design. What tells a lineage its taste was good? The success of its offspring **against reality**. The
reality-check that stops collapse (grounding, §5) and the fitness signal that drives the evolving taste
turn out to be the *same thing*, seen from two sides.
## 7. The central danger: fitness is not truth
Introducing selection introduces selection's classic hazard, and it is severe enough to sink the whole
scheme if ignored. Evolution optimises, without mercy or foresight, for exactly what you *measure*
never for what you *meant*. (Economists and ML engineers know this as **Goodhart's law** and
*specification gaming*.) Get the fitness measure slightly wrong and the lineage will exploit the gap
with more ingenuity than any designed rule.
For a *knowledge* lineage there is a specific and nasty version. For ideas, the natural measure of
"fitness" is **how well they spread**, and a false-but-persuasive idea spreads beautifully. Human
intellectual culture is full of highly transmissible falsehoods; confident nonsense out-competes hedged
accuracy in almost every human forum. Turn Darwinian selection loose on models without care and it will
breed a lineage optimised for *persuasiveness* — fluent, compelling, and wrong. That is model collapse
with an optimiser behind it, actively seeking the cliff.
Only one thing makes fitness track truth rather than appeal: **being judged against a reality that can
say no.** Fitness must be predictive success under *intervention* — did the model's knowledge correctly
anticipate what the world would do when acted upon — and not approval, fluency, or a benchmark score,
each of which can be gamed. This is why the reality-check is load-bearing twice over: it is both the
anchor that stops passive collapse *and* the only thing that keeps the evolving taste honest.
The second danger is **convergence**, and beating it takes work at two separate levels, because
selection can only preserve variety that already exists — the variety must first be *supplied* and then
*kept*.
- **Supply.** A lineage that learns only from an accredited elite has a monoculture for a source: the
"best" experts are, almost by definition, the ones who won the consensus, so the incoming variation
is narrow from the start. The society must therefore learn, deliberately and from the beginning, from
the **outliers and the heterodox** as well as the credentialed — not out of fairness, but because in
evolutionary terms diverse founders are the raw material without which nothing downstream can adapt.
- **Preserve.** Even given varied input, plain fitness-*maximising* selection converges — it drives
every lineage toward the single current best and fixes it, extinguishing the rare specialists. The
fix is well established: **quality-diversity** selection, which rewards being *good* and being
*different* at once (novelty search and MAP-Elites — Lehman & Stanley, 2011; Mouret & Clune, 2015),
keeping complementary specialists alive rather than collapsing onto the champion. In our simulations
this is decisive: greedy "keep-the-best" selection collapses a population's diversity almost at once
and gets stuck at a mediocre answer, while quality-diversity selection keeps the specialists that
sexual recombination then needs as parents.
The two levels meet at reproduction. Multi-parent recombination (§5) is the *vehicle* by which the
diversity this selection preserves actually enters the next generation: an offspring drawn from
complementary parents inherits the standing variation the selector kept alive, recombined into one new
model. Supply the variety from the human side; preserve it on the selection side; recombine it into
each generation on the reproduction side. Remove any of the three and the lineage converges on its own
first guess.
## 8. A society needs institutions, not just specialists
One requirement is easy to overlook and fatal to omit. The easy part of a society is specialisation.
The *hard* part — which human civilisation took millennia to build — is the set of **institutions that
let fallible specialists combine without each re-verifying everything**: reputation, replication,
credentials, and above all **peer review**. These are error-correction protocols, and they exist
because a group of unreliable specialists left to reinforce one another is *more* wrong than any member
alone.
This is precisely where current multi-agent AI fails: set several models to confer and they tend to
agree sycophantically and confabulate in committee, because they have all the specialisation and none
of the institutions. A multigenerational society must specify not only how models learn, reproduce, and
are selected, but how they *check* one another — how a claim is challenged and a mistaken model loses
standing *before* its error is recombined into offspring and inherited. Peer review is itself a
reality-check of the kind §7 demands — an institutional stand-in for reality's "no," to be used where
direct intervention is slow or costly.
## 9. The lineage must stay open to reality
A society of models, however many generations deep, shares one hard limit: it has only ever *read*.
Its whole inheritance is a record of things that were said. In the vocabulary of causal reasoning
(Pearl, 2009), it lives on the bottom rung of the **ladder of causation** — observation — and no amount
of observation reaches *intervention*. Watching underdetermines doing; correlation does not contain
causation, at any scale.
Only intervention — reaching out and changing the world to see what happens — climbs the ladder, and a
language model cannot intervene. This is what humans and their instruments supply, and the contribution
is not "truth" but **constraint**: reality's unique gift is that it can say **no**. Text offers only
more opinion; an experiment delivers a refusal no consensus can overturn. As §§67 argued, that refusal
does double duty — it is both the anchor that prevents collapse and the fitness signal that lets the
lineage's evolving taste select for truth rather than persuasion.
Two honest riders. First, the human reality-signal is *dirty*: people supply results warped by
publication bias, incentive, and occasional fraud — which is exactly why the error-correcting
institutions of §8 must sit at the humanmachine boundary, screening the signal before it selects.
Second, humans are the *current* supplier of intervention, but the actuator half is being automated
(autonomous laboratories already close the designbuildtest loop). What looks durable in the human
role is therefore not the hands but the **choice of what to test and which refusals matter** — the
part of the fitness function that encodes *what is worth persisting*, as opposed to what merely *can*
persist. We flag, without resolving, that a partnership stays mutual only while both sides supply
something the other cannot.
## 10. Why it is cheap
A practical fact turns this from thought experiment into buildable proposal: **the architecture almost
never re-pays for the one genuinely expensive thing in AI — pre-training.** (The single exception,
periodically re-minting the base, is §11, and it is rare enough to be an amortised footnote.)
Training a foundation model from scratch consumes trillions of words and a fortune in compute. This
design does none of that per generation. Every model is *born* from an existing open-weight model that
already paid that cost; specialising one is a small patch trained in hours on a single consumer GPU;
running the society is ordinary inference; and reproducing — recombining parents into a child — is, in
the model-merging case, cheaper still, because it can be done directly on the weights with no retraining
at all (Akiba et al., 2024). Selection does cost more — you must run *populations* and discard the
unfit — but that is a multiplier over an already-cheap unit, not over a foundation-model budget.
The economics work only with **open-weight** models, for reasons practical and legal at once: you must
be free to inspect, modify, and redistribute the weights, and most proprietary licences forbid using a
model's outputs to train another — which is exactly what reproduction here does. This is not ideology
bolted on; it is a structural constraint, and a democratising one, since it puts the whole architecture
within reach of a single laboratory.
## 11. Can it grow forever? Consolidating knowledge back into the base
One question the design has assumed away: can the lineage accumulate *without end*? The individual is
bounded, and that is the clock. But the lineage seemed unbounded — each generation simply starts a
little ahead. Look closer and a second budget also fills.
Every new model is a pristine base plus an inherited **soft** delta — the acquired knowledge carried in
added patches rather than baked into the frozen core (§3). That soft delta is what makes the lineage
multigenerational; it is also what cannot grow forever cheaply. Stacked patches are not free: they slow
inference, and past some depth the accumulated delta is better *consolidated* than carried. The lineage,
too, matures.
The fix is the same operation, one level up. When a lineage's acquired knowledge has proven stable
across enough generations, **re-mint the base**: distil the accumulated soft inheritance into the
*weights* of a fresh foundation-scale model — a new base born already *natively knowing* what took many
generations to acquire in patches. The soft budget resets; the next epoch begins from a richer floor.
What was hard-won and *learned* becomes cheap and *innate*.
This has a precise name, and it is not Lamarck's. Knowledge that is acquired and re-learned every
generation, and — once reliably present for long enough — becomes part of the innate endowment so that
it need no longer be re-learned, is the **Baldwin effect** (Baldwin, 1896; and its clean computational
demonstration, Hinton & Nowlan, 1987). It is the valve between the two substrates: the soft, learned
patches, and the hard base weights every model is born with.
Three honest riders, because re-minting is the most consequential step in the scheme:
- **Cost.** This is the one step that re-pays part of the pre-training bill, breaking §10's cheapness
*locally*. It is bearable only because it is *rare*, amortised over many cheap generations, and is
continued training from the lineage's own rich outputs rather than a de-novo run.
- **Irreversibility.** Until now, one thing was always recoverable — the original pristine base, whose
lost tails could be restored just by reloading the file. Bake the current lineage into new immutable
weights and that escape hatch closes: if the lineage had been quietly collapsing, re-minting *fixes
the collapse in place* and discards the one uncollapsed reference that could have diagnosed it. In our
minimal models this is exactly what happens, and a cheap safeguard prevents it: **re-mint only while
the lineage is demonstrably diverse and healthy**, never as a rescue for a line already drifting. It
is the sharpest instance of the human seat of §9 — choosing what no future generation will think to
question.
- **Speciation.** A re-minting is a founder event. Different laboratories, re-basing on different
criteria, will mint divergent bases; the lineage branches. This is not a defect but *adaptive
radiation*, and it is exactly what open weights make possible. The society grows not as one heavy
trunk but as a branching tree of bases.
So the answer to "can it grow forever?" is **yes — but only because it forgets and consolidates at
every level, including the base.** Nothing is retained without bound anywhere; unbounded growth of
*capability* is bought by *bounded* storage plus periodic consolidation.
## 12. One process, four timescales
Step back and the parts resolve into a single idea running at four nested speeds. The **vertical**
motion is transmission — the selective passing-down of hard-won knowledge:
1. **Within one model, over a working life:** experience is consolidated from fast, episodic memory
into slow, durable weights, without catastrophic loss.
2. **Between generations, at maturity:** mature models reproduce — recombined into a fresh one.
3. **Across many generations:** each generation inherits the compressed achievements of the last and
builds on them.
4. **Across epochs:** a proven lineage's accumulated soft inheritance is consolidated into the weights
of a re-minted base, becoming innate.
The first and last are the *same operation at opposite ends of the scale* — a fast/soft store
consolidating into a slow/hard one — one running overnight inside a single model, the other across an
epoch inside a whole society. The **horizontal** motion is selection — Darwinian selection acting across
the population at each timescale, on the policies that govern what gets transmitted, with reality as the
fitness function and diversity-preservation keeping the specialists alive.
The same three rules govern all of it: **reproduce by recombining, not by copying, or you decay;
preserve the disagreements and the surprises, or you converge; and anchor fitness to a reality that can
refute, or you evolve toward what is merely convincing.**
## 13. What we built, what we found, and what is still open
The previous drafts of this paper promised a "companion paper" that *would* make this concrete. That
work now exists — mostly as a set of **minimal, laptop-reproducible models**, with a first bridge to
**real language models** (a LoRA-merge prototype, up to 7B on a GPU cluster) — and it is worth stating
plainly what it does and does not show. (A separate results document gives the numbers; here is the
shape.)
**What we built and found.**
- *An exact account of collapse.* Because generational training is the WrightFisher drift process, we
can check a simulator against century-old closed-form formulas, and it matches them to a fraction of
a percent. Collapse is not argued by analogy; it is derived.
- *The cheap-grounding result, and its limit.* A few percent of verified real data holds on to most of
a lineage's diversity indefinitely — but not the deepest tail, which needs recombination. This is
what makes a continually-learning society economically plausible rather than a data-hungry fantasy.
- *"Merge, don't average."* Combining several teachers by *averaging* their outputs — the obvious thing,
and what a "model soup" does — mathematically cancels the benefit of having several teachers. A
*merge* that keeps each item's strongest source realises it. Most current multi-model setups get this
wrong by default.
- *Collapse and its cure in real trained networks, and on real images.* We reproduced the same effects
in small recurrent and feed-forward networks and in a generator of handwritten digits (MNIST), where
a model trained on its own output collapses to a single blurred digit while a little grounding keeps
all the styles alive. An honest wrinkle we had to report: real neural networks *smooth*, so the naive
diversity metric misleads, and the right measure is distance-from-truth.
- *Sex that beats the parents, and when it doesn't.* In evolutionary simulations, recombining
complementary specialist models produces a model fitter than any parent (the FisherMuller effect),
climbing toward the best-possible combination as more, more-diverse parents are added — while
averaging and best-single-parent plateau below. On *entangled* problems, blind recombination instead
produces below-parent offspring (outbreeding depression) — and *directed* recombination (choose mates,
screen offspring, unbounded parents) reliably fixes it. This is the concrete evidence for the paper's
central reframing.
- *The recombination claims, in real language models — with a sharp condition.* Merging LoRA-specialised
Qwen models (up to 7B on a GPU cluster) produces a generalist that beats every specialist parent
(FisherMuller, for real); and keeping parents intact and *routing*, or *breeding and screening*
offspring, beats the naive average — but *only when the task leaves headroom*. On easy tasks a strong
model's plain average is already at the ceiling and the refinements add nothing; on hard tasks the
average dilutes a specialist below even the best single parent, and the union-preserving operators win
clearly. The practical rule is exact: these tricks pay off in proportion to how far the naive average
is from the best attainable. This is a prototype (three task families, one seed), so we read it as
signs, not magnitudes; the *whole grounded society* on a language model remains the open step.
- *The whole society, and why every part is needed.* In a population evolving on a "reality" landscape,
the full system — grounding + sexual recombination + preserved diversity — climbs to the top while
keeping its specialists. Remove *grounding* and it collapses into a confident, wrong consensus (a
direct analogue of training on the internet's growing crowd of AI-generated text); remove *sex* and it
gets stuck; remove *diversity* and it converges too fast to a worse answer. Each removal fails
differently; only the whole system climbs. This is the closest thing we have to a test of the actual
thesis, rather than of the borrowed scaffolding around it.
**What is borrowed, and what is ours.** We want to be careful here, because one part of the story is no
longer novel. The reading of *model collapse as genetic drift* — the core diagnosis — has since been
derived independently and more rigorously than we had (Riis, 2026), and we cite it as such; we do not
claim it. What we do claim is the **cure and its assembly**: grounding as immigration from a fixed
reality (which yields the cheap-grounding result a closed, self-consuming loop cannot); recombination
reframed as **sexual reproduction**, with the "merge-don't-average" law, the FisherMuller "offspring
exceed parents" result, the outbreeding-depression limit, and directed sex as the engineered advantage
over biological sex; the observation that real trained networks deviate from the neutral drift model in
a characterisable, architecture-specific way; and the integrated society in which grounding, sex, and
diversity together produce a climbing lineage. In one sentence: the field increasingly agrees on the
*disease*; our contribution is a **control theory for the cure**.
**What is still open — honestly.** The old hole (what to select) we fill in kind: don't design the
selector, evolve it. But the hole has *moved*, not closed, and the new one is harder: **the fitness
function** — what reality-anchored measure selects for *truth* without also selecting for *persuasion*,
given that in our own species the two have been at war for the whole history of ideas. Alongside it:
the **institutions** that let contemporaries correct one another before error is inherited (§8), which
we do not solve; and the **calibration** of everything the results left as knobs — how many parents,
how complementary, at what ratio of inherited-to-real data, and how healthy a lineage must be before
its knowledge is safe to make irreversibly innate. These are, at least, *measurable* — which is the
difference between an open problem and a hole. And the largest gap of all: the *recombination* claims
now hold in real language models, but the *society* — the grounded, diversity-preserving, continually
reproducing loop — does not yet. The real test is to build that whole system out of actual open-weight
language models, and see whether all the signs survive contact with a system too big to write down.
The operators, checked; the living society, next.
---
## Selected references
- Akiba, T., Shing, M., Tang, Y., Sun, Q., & Ha, D. (2024). Evolutionary optimization of model merging recipes. *Nature Machine Intelligence.* (See also Sakana AI's M2N2, "Model Merging of Natural Niches.")
- Baldwin, J. M. (1896). A new factor in evolution. *The American Naturalist.*
- Boyd, R., & Richerson, P. J. (1985). *Culture and the Evolutionary Process.*
- Campbell, D. T. (1974). Evolutionary epistemology. In *The Philosophy of Karl Popper.*
- Fisher, R. A. (1930). *The Genetical Theory of Natural Selection.*
- French, R. M. (1999). Catastrophic forgetting in connectionist networks. *Trends in Cognitive Sciences.*
- Hinton, G. E., & Nowlan, S. J. (1987). How learning can guide evolution. *Complex Systems.*
- Hinton, G., Vinyals, O., & Dean, J. (2015). Distilling the knowledge in a neural network. *arXiv:1503.02531.*
- Hu, E. J., et al. (2021). LoRA: low-rank adaptation of large language models. *arXiv:2106.09685.*
- Hull, D. L. (1988). *Science as a Process.*
- Kauffman, S. A., & Levin, S. (1987). Towards a general theory of adaptive walks on rugged landscapes. *Journal of Theoretical Biology.* (The NK model.)
- Lehman, J., & Stanley, K. O. (2011). Abandoning objectives: evolution through the search for novelty alone. *Evolutionary Computation.*
- Mallya, A., & Lazebnik, S. (2018). PackNet: adding multiple tasks to a single network by iterative pruning. *CVPR.*
- McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Why there are complementary learning systems in the hippocampus and neocortex. *Psychological Review.*
- McCloskey, M., & Cohen, N. J. (1989). Catastrophic interference in connectionist networks. *Psychology of Learning and Motivation.*
- Minsky, M. (1986). *The Society of Mind.*
- Mouret, J.-B., & Clune, J. (2015). Illuminating search spaces by mapping elites (MAP-Elites). *arXiv:1504.04909.*
- Muller, H. J. (1932). Some genetic aspects of sex. *The American Naturalist.* (The advantage of recombination.)
- Muller, H. J. (1964). The relation of recombination to mutational advance. *Mutation Research.* (Muller's ratchet.)
- Pearl, J. (2009). *Causality: Models, Reasoning, and Inference* (2nd ed.).
- Popper, K. (1959). *The Logic of Scientific Discovery.*
- Riis, S. (2026). Drift and selection in LLM text ecosystems. *arXiv:2604.08554.*
- Rusu, A. A., et al. (2016). Progressive neural networks. *arXiv:1606.04671.*
- Shumailov, I., et al. (2024). AI models collapse when trained on recursively generated data. *Nature.*
- Wortsman, M., et al. (2022). Model soups: averaging weights of multiple fine-tuned models. *arXiv:2203.05482.*
- Wright, S. (1931). Evolution in Mendelian populations. *Genetics.*
*Literatures the next version should still engage: multi-agent LLM societies (to mark the departure); population-based training and open-ended evolution; tacit knowledge (Polanyi) and human capital (Becker).*