SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution

Researchers introduced SkillRise, a unified reinforcement-learning framework for LLM agents to extract, refine, and reuse skills across related tasks.

The single policy alternates between solving tasks and updating a shared skill document, enabling cross-task skill transfer without a multi-stage pipeline.

On ALFWorld, WebShop, and ScienceWorld, SkillRise beat prior methods by 2.3 to 8.5 percentage points in Pass@1.

The approach remained effective across repeated attempts and longer task sequences, while reducing runtime overhead.