SkillRise trains LLM agents to build reusable skills across task sequences
A new reinforcement learning framework teaches language model agents to extract and refine skills across progressively harder tasks, outperforming baselines by 2.3 to 8.5 percentage points on ALFWorld, WebShop, and ScienceWorld.
SkillRise is a reinforcement learning framework that teaches LLM agents to extract and reuse skills across related tasks. Standard agentic RL treats each task as an independent episode, while existing skill-learning pipelines split extraction, retrieval, and execution into separate stages. SkillRise uses a single policy that alternates between solving the current task and curating an evolving skill document passed directly to the next task in a progressively harder sequence. Credit assignment is decoupled: solving is supervised by the current task outcome, while curation is supervised by discounted downstream outcomes.
Experiments on ALFWorld, WebShop, and ScienceWorld show SkillRise achieves the strongest Pass@1 performance among compared methods, with gains over the strongest baseline ranging from 2.3 to 8.5 percentage points. The learned curation policy remains effective when applied to repeated attempts on the same task, even though it was trained across distinct tasks. Performance improves with longer sequences of related tasks at test time, even when each task is attempted only once—evidence that the framework reuses transferable skills rather than benefiting from repeated sampling. SkillRise retains strong performance while substantially reducing the runtime overhead of multi-stage skill learning pipelines. The preprint was posted to arXiv on July 30, 2026.
