Skill Adapter Study (GPT-5 Mini executor) - SkillsBench Success Rate: leaderboard
Benchmarks and budgets: ALFWorld 420 test tasks / 50 steps (six task types); WebShop 1,500 tasks / 15 steps; SkillsBench 210 tasks / 30 steps (Experimental Setup). Executors: Qwen3-8B (ALFWorld, WebShop), GPT-5-mini (SkillsBench); frozen; shared by all methods within a benchmark (Experimental Setup). Skill pools: GPT-5.2-Codex-extracted skills from clustered AgentTraj-L trajectories (ALFWorld, WebShop); merged benchmark-provided skills (SkillsBench) (Experimental Setup). Retrieval (Skill Adapter): BGE-M3 embeddings; 3 topics, 20 skills, expansion capped at 30 candidates; router selects at most 8 skills (Experimental Setup). Metrics: ALFWorld per-type and overall SR, mean time; WebShop task score, SR, mean time; SkillsBench SR, mean reward, mean time (Experimental Setup, Metrics). Runs: Seeds 42, 43, 44; main tables report the mean of three runs (Experimental Setup). Decoding, thinking mode (Qwen3-8B), reasoning effort and snapshot (GPT-5-mini): not stated (not stated). Board structure: Each benchmark has a single executor identity, so every admissible board ranks configurations of one model; under the per-board harness fold these boards reduce to one model row and carry no cross-model ranking signal. Materialize only if single-identity configuration boards are kept. (reviewer note). This board: 210 SkillsBench test tasks; Mean over three runs (seeds 42, 43, 44). SkillsBench (Li et al. 2026b) test set as used by the paper (210 tasks); version not stated. Metric: SkillsBench task success rate (%) Higher is better.
Metric: SkillsBench task success rate (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 1 model tracked. Updated 14 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | GPT-5 Mini | Skill Adapter run: Backbone + BM25 | 13.33 | #354 |
Interactive version: theaggregate.ai/benchmark?slug=skill-adapter-study-gpt-5-mini-executor-skillsbench-success-rate · How It Works · Data refreshed daily, snapshot 2026-11-16.