inference-sh/skillos-alfworld-eval-arms

SkillOS reproduction — ALFWorld paired eval rollouts

Per-game outcomes for every ALFWorld evaluation arm in an independent reproduction of SkillOS. This is the evidence behind every number in the report: with these files you can recompute each significance test yourself, on a laptop, without a GPU or an API key.

Each eval-*/ directory is one sweep; each ckptN.jsonl is one arm; each line is one held-out ALFWorld game. Arms within a sweep are paired by gamefile, which is what makes McNemar the right test — the same 140 valid_seen games in the same order for every arm.

git clone https://github.com/belt-sh/skillos && cd skillos
python -m scripts.compare_eval_arms \
  --arm no_memory=eval-pathbv4/no_memory.jsonl \
  --arm ckpt30=eval-verl-gigpo-real/ckpt30.jsonl
# -> 40.7% vs 33.6%, +7.1pp, McNemar p=0.0987

Directories

directory	what it is	arms
eval-pathbv4	the canonical 33.6% no-memory baseline (no_memory.jsonl) reused by every 8B sweep	7
eval-verl-gigpo-real	verl/GiGPO run, every-5 sweep	12
eval-fft, eval-fft-seed2, eval-fft-seed3	TRL full fine-tune, 3 seeds	8–12 each
eval-v8	TRL LoRA r=32 run (weights no longer exist; these rollouts are all that remains)	6
eval-fft-natural, eval-fft-curriculum	the paper's two grouping ablations — both null	12 / 9
eval-transfer-32b*	same games, Qwen3-32B executor, vs a 49.3% no-memory baseline	3–12 each
eval-reasoning-to-alfworld	cross-domain: a DeepMath-trained curator evaluated on ALFWorld	12
eval-decode	executor decode-parameter sweep (null)	4

Two things worth knowing

Pool the arms before you believe any of them. Across the five 8B sweeps here there are 50 arms tested against one baseline. Family-wide Bonferroni sets the bar at p < 0.001; the best arm is p = 0.0026. No same-executor lift survives correction. The per-arm p-values in these files are uncorrected.

eval-reasoning-to-alfworld is the strongest signal in the set, and it is negative. The paper reports +13.3pp on ALFWorld from a reasoning-trained curator; these rollouts give −14 to −18pp (p ≤ 0.0005) at every checkpoint past step 40.

GPQA is not here

The reproduction also evaluated GPQA-Diamond, a gated dataset whose access condition forbids publishing problem text, options, answers or model responses. No GPQA rollout is included in this repo, and none ever will be — only aggregate accuracies appear in the report. Reproducing that column requires your own access. See data/README.md in the code repo.

AIME rollouts are also excluded here; this dataset is ALFWorld only.

Provenance

• Code, full report and every divergence from the paper: https://github.com/belt-sh/skillos
• Findings write-up with figures: docs/repro_report.md
• Paired eval rollouts, so every number here can be recomputed without a GPU: inference-sh/skillos-alfworld-eval-arms
• Paper: SkillOS: Learning Skill Curation for Self-Evolving Agents (Ouyang et al., 2026)

This is an independent reproduction. It is not affiliated with or endorsed by the paper's authors.

License: Apache 2.0, inherited from Qwen/Qwen3-8B.