Giorgio Gilestro
5a23ddaf2a
llm_speciation (new kind; src/llm/speciation.py): E13 in LLM weights. LoRA children share the frozen base's coordinates, so merge failure is functional by construction. CONFLICT (ambiguous sort prompts learned under opposite conventions — the BDM structure): function-specific hybrid breakdown — merged coherence 0.02-0.08 falls below BOTH parents (~0.2) on the conflicted function; and in the de-confounded `add` design (private budget fixed, conflict added on top; 3 seeds after a single-seed pilot showed one anomalous point) the merge's private-family accuracy shows NO trend with conflict — the damage is surgical, not global. DURATION (over-trained disjoint specialists, 1->12 epochs): the merge improves (0.84->0.94) and stays above the best parent — the MLP "no emergent isolation" null generalises; relevant to the expert-training-duration report (2607.11997), with the epistasis prediction left to the decisive experiment. Multi-seed firm-up (seeds threaded into specialist caches; `seeds:` list support in the runner; fixed test sets): all three recombination claims hold with CIs — merges beat every specialist (5 seeds, ties 0.647±0.027 > best spec 0.592±0.009; worst-family 0.28 vs <=0.16); union 0.274±0.026 > fusion 0.174±0.102 on hard (3 seeds); directed 0.221±0.026 > soup. NEW finding: fusion is seed-FRAGILE where headroom exists (CI ±0.10) while routing/directed selection are stable (±0.026) — the union/selection operators win on reliability, not just mean. Figures (llm_speciation 3-panel; llm_seeds 3-panel with 95% CI), READMEs, +1 convention test (150 green), make llm-speciation / llm-seeds targets. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BkRLcc18rwT2Lysu6PbG7v
3.7 KiB
LLM-tier model speciation — conflict provokes isolation; duration alone does not
E13 carried into language-model weights (0.5B Qwen, LoRA children of one frozen base — which shares
its coordinate system with both children, so there is no permutation/rescaling ambiguity by
construction: every merge failure here is functional). Two knobs, pre-registered readings in the
configs; figure llm_speciation.png (3 panels; panel B from results/llm_speciation_add/).
Design. Child A: private family strings; child B: private family arith; shared ambiguous
convention prompts ("Sort the list […]" — no direction stated) answered ascending by A and
descending by B: each convention harmless alone, contradictory jointly (the Bateson–Dobzhansky–
Muller structure). 50/50 soup merge; exact-match verifier; fixed test sets. Coherence of a model =
max(accuracy under asc grading, under desc grading) on the shared prompts — a coherent parent scores
under its own convention; a hybrid mixing conventions scores low under both (the μ(S)/2 floor made
operational).
Finding 1 — function-specific hybrid breakdown (the conflict knob; panel A)
Once the conventions are trained (conflict_frac ≥ 0.25), each parent performs under its own
convention (~0.20–0.25 — low absolute values: sorting long lists is hard for 0.5B) while the merge's
coherence sits at 0.02–0.08, below BOTH parents under either grading — the hybrid loses precisely
the conflicted function. (At frac = 0 no one has seen the ambiguous prompts; that point is a
no-conflict baseline, not part of the cliff.)
Finding 2 — the damage does not spread (the de-confounded add design; panel B)
In the original ("replace") sweep, higher conflict fraction mechanically means less private-family
training, so the merge's private-family decline is confounded. The add design
(llm_speciation_add, 3 seeds) holds each child's private training fixed and adds conflict data on
top: the merge's private-family accuracy then shows no trend with conflict (0.74–0.88, tracking
parent A's 0.82–0.87 within seed noise at every level). Conflict damage is localised to the
conflicted function; it does not corrupt the disjoint skills — at this scale, hybrid breakdown is
surgical, not global. Honest rider: 0.5B soup merges carry large intrinsic seed variance even at
zero conflict (sd up to 0.28) — the same averaging-fragility seen in llm_moe_hard_seeds.
Finding 3 — the duration null: over-specialisation does not erode mergeability (panel C)
Pure disjoint specialists over-trained from 1 to 12 epochs (no shared data at all): the merged model
improves (0.84 → 0.94 mean-private) and stays above the best parent at every duration. The MLP
tier's "no emergent isolation" null (speciation_real_emergent) generalises to LLM weights in
this regime — relevant to the report that averaging prefers under-trained experts (arXiv:2607.11997):
in our disjoint-family setting, no such over-training penalty appears; the theory's prediction is that
their effect should trace to conflicting conventions on shared circuitry, which the
epistasis_predicts experiment (work order) will test directly.
Speciation across all three tiers now reads: analytic (E12: cliff, epistasis-dependence,
snowball) → MLP (E13: functional residual survives the full symmetry group; no emergent isolation) →
LLM (this run: function-specific hybrid breakdown under conflict; no isolation from duration or
specialisation alone). Isolation must be provoked by functional conflict at every tier tested.
Falsifiers (not triggered): merge coherence matching the parents (no breakdown), or merged
private-family accuracy declining with conflict in the add design (global corruption).