Giorgio Gilestro
84124de143
Clarity pass over the main text (36-item audit), Discussion rewrite and cut, acknowledgements, Souly et al. as ref 62, lettered SI panels, model section moved under Results; plus the untracked curriculum/society/compose/smol configs, runners, figures, stats and tests that the SI already cites. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Y64o8FKP7rCuXzC48pxpMm |
||
|---|---|---|
| .. | ||
| s2 | ||
| s3 | ||
| manifest.json | ||
| partial_isolated_s1.parquet | ||
| partial_society_s1.parquet | ||
| README.md | ||
| resolved_config.yaml | ||
llm_curriculum_v5_decor — a curriculum that decouples partner complementarity from generation
Config configs/llm/curriculum_v5_decor.yaml. Same population as v5 (three lineages, six real-dataset
families, Qwen2.5-1.5B, rank-16 adapters continued from the parent) but every lineage starts with
mnli, then diverges maximally, then converges, so mean pairwise complementarity by generation is
0.00, 0.67, 0.70, 0.58, 0.33, 0.00 (Latin square: 1.00, 1.00, 0.80, 0.67, 0.33, 0.00). Arms:
isolated (never merge) and society with allow_veto: true (declinable merge). Pre-registered in
tasks/prereg-llm-society-v4.md §8h before running. Seed 1 local; seeds 2–3 on CX3
(hpc/llm_curriculum_controls.pbs). Layout: results.parquet (seed 1), s2/, s3/.
Why: in the Latin square, complementarity and generation number are collinear, so the veto's acceptance curve (Fig. 4B, Spearman ρ = −0.57 with complementarity) could equally be a response to adapter age. Pooling both curricula breaks the collinearity.
Result (figures/stats_llm_curriculum.py)
Fraction of proposed merges declined (mean of 3 seeds) against complementarity:
| generation | Latin: declined | Latin: compl. | decor: declined | decor: compl. |
|---|---|---|---|---|
| 0 | 0.44 | 1.00 | 0.44 | 0.00 |
| 1 | 0.56 | 1.00 | 0.44 | 0.67 |
| 2 | 0.44 | 0.80 | 0.67 | 0.70 |
| 3 | 0.78 | 0.67 | 0.67 | 0.58 |
| 4 | 0.67 | 0.33 | 0.44 | 0.33 |
| 5 | 1.00 | 0.00 | 0.89 | 0.00 |
Pooled test (2 curricula × 6 generations × 3 seeds = 36 points):
| quantity | value |
|---|---|
| partial Spearman ρ(declined, complementarity | generation) | −0.067, seed-clustered 95% CI (−0.211, +0.088) |
| partial Spearman ρ(declined, generation | complementarity) | +0.31 |
| raw ρ(declined, complementarity) | −0.314 |
| raw ρ(declined, generation) | +0.458 |
Final accuracy (best lineage, all six families, generation 5):
| arm | s1 | s2 | s3 | mean |
|---|---|---|---|---|
| decor, declinable merge | 0.789 | 0.792 | 0.789 | 0.790 |
| decor, never merge | 0.786 | 0.792 | 0.792 | 0.790 |
Pre-registered reading (§8h): the modifier hypothesis (partial ρ with complementarity < 0, CI excluding 0) is not supported; the adapter-age hypothesis (partial ρ with complementarity ≈ 0, with generation > 0) fits. The Latin-square correlation of −0.57 was carried by generation. What rises with generation in both curricula and is not separated here: the adapters' training age, the number of skills each holds, and the presence of the two format-destroyer families (boolq, winogrande), which every lineage carries by the late generations in both schedules. The declinable merge again finished level with never merging (−0.000 ± 0.003 over seeds).
Falsifier status: the manuscript's reading of the acceptance curve as a recombination modifier tracking complementarity (the reduction principle) is withdrawn; the curve tracks generation.