SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution
SkillRise trains an agent to carry a living skill document across related tasks, improving single-try performance.
The framework alternates between solving each task and updating reusable skills for the next one. Its credit assignment scores task-solving on the current outcome, while skill curation is judged by discounted downstream results. In tests on ALFWorld, WebShop, and ScienceWorld, it beat the strongest baseline by 2.3 to 8.5 Pass@1 points. The authors also report better performance on longer related task sequences and lower runtime overhead than multi-stage skill-learning pipelines. HF Daily Papers' note
The framework alternates between solving each task and updating reusable skills for the next one. Its credit assignment scores task-solving on the current outcome, while skill curation is judged by discounted downstream results. In tests on ALFWorld, WebShop, and ScienceWorld, it beat the strongest baseline by 2.3 to 8.5 Pass@1 points. The authors also report better performance on longer related task sequences and lower runtime overhead than multi-stage skill-learning pipelines. HF Daily Papers' note
score 5