SkillForge: A skill library that continuously verifies instead of only appending¶

Explicit skill use, posterior update and revision mini-suite.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	AMAP, Alibaba Group (first-listed first-author affiliation).
First public release	2026-08-25 (arXiv v1)
Author code	No public author code found as of 2026-08-26.
Local adapter or method key	`skillforge`
Local reproduction code	`src/auto_research/agent_research/latest_20260826.py`

Paper method and evidence¶

Appending extracted skills forever can retain erroneous or stale entries. Include explicit skill invocation in RL, induce candidates from successful, failed and contrastive traces, and use environment evidence to activate, revise or deduplicate.

flowchart LR
  A[Explicitly invoke skill] --> B[Collect contrasting outcomes]
  B[Collect contrasting outcomes] --> C[Update posterior and revise library]

Original paper figure¶

The original figure shows evidence-driven skill evolution. From the original paper; copyright remains with the authors.

Core equation¶

\[ \max_{\theta,\mathcal B}\;\mathbb E_{d\sim\mathcal D,\tau\sim\pi_\theta(\cdot\mid d,\mathcal B)}[r(\tau)], \]

\[ \mathcal S_{ret}=\operatorname{TopK}_{s\in\mathcal B}\cos(e_d,e_s),\quad a_t=(a_t^{env},c_t). \]

Reported results¶

Without cold-start initialization, paper mean +6.3% versus SkillRL. Qwen3-4B ALFWorld/WebShop/AppWorld 87.9 / 84.0 / 44.6. No industrial A/B.

Local reproduction¶

PlanBench mini 120 episodes: 12 new skills, 108 reuses, joint success 1.0000, mean cost 0.6200. Metrics, batch index. Structured plans stand in for skills; no LLM induction, three external benchmarks or RL training.

auto-research agent-eval --method skillforge --benchmark planbench-mini --episodes 120 --seed 42