SkillRL (NeurIPS 2026): 7B model beats GPT-4o by 41% via recursive skill evolution

cihangxie · x · 2026-09-25

SkillRL (arXiv:2602.08234), accepted to NeurIPS 2026, tackles LLM agents' failure to accumulate skills from experience. It proposes recursive skill-augmented RL: experience-based distillation builds a hierarchical skill library (SkillBank), an adaptive retrieval strategy covers general and task-specific heuristics, and a recursive evolution mechanism lets the skill library co-evolve with the policy during RL — cutting token footprint while boosting reasoning utility.

On ALFWorld, WebShop, and seven search-augmented tasks it sets SOTA: a 7B model outperforms GPT-4o by 41%, with 20% fewer training tokens and 33% faster convergence. Paper, data, and models are open-sourced on arXiv, HuggingFace, and GitHub.

Original post →

More from coding & agent

coding & agent channel →