Researchers from Hugging Face have published a paper introducing SkillRise, a novel reinforcement learning (RL) framework designed to enable large language model (LLM) agents to learn and transfer skills across related tasks. The work addresses a key limitation in current agentic RL, where tasks are typically treated as independent episodes, preventing agents from reusing knowledge.
SkillRise organizes related task instances into progressively challenging sequences. A single policy alternates between solving the current task and curating an evolving skill document that is passed to subsequent tasks. The framework employs decoupled credit assignment: task solving is supervised by the immediate task outcome, while curation is supervised by discounted downstream outcomes. This allows the agent to learn which skills are worth preserving for future tasks.
The researchers evaluated SkillRise on three benchmarks: ALFWorld, WebShop, and ScienceWorld. SkillRise achieved the strongest Pass@1 performance among compared methods, with gains over the strongest baseline ranging from 2.3 to 8.5 percentage points. Notably, the learned curation policy remained effective even when applied to repeated attempts on the same task, despite being trained across distinct tasks.
Further analysis revealed a scaling property: performance improved with longer sequences of related tasks, even when each task was attempted only once. This suggests that SkillRise reuses transferable skills across tasks rather than benefiting from repeated sampling of the same task. Additionally, SkillRise substantially reduced the runtime overhead compared to multi-stage skill learning pipelines, while retaining strong performance.
The authors conclude that SkillRise provides a simple and efficient training paradigm for LLM agents to extract, refine, and reuse transferable skills across tasks, opening new avenues for cross-task generalization in agentic AI.