SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution

SkillRise introduces a unified reinforcement-learning framework for agents that learn reusable skills across related but distinct tasks. It arranges task instances into progressively harder sequences and uses one policy to alternate between solving the current task and curating an evolving skill document for the next task. Credit assignment is decoupled: solving is supervised by the current outcome, while curation receives discounted downstream rewards. On ALFWorld, WebShop, and ScienceWorld, SkillRise reports the best Pass@1 results among compared methods, improving over the strongest baseline by 2.3 to 8.5 percentage points. The paper also reports test-time scaling across task sequences and lower runtime overhead than multi-stage skill-learning pipelines.