SkillRise Enables Cross-Task Skill Evolution for LLM Agents.

Large language model (LLM) agents frequently encounter tasks that, while distinct, share underlying solution patterns or reusable skills. Current agentic reinforcement learning often treats each task as an isolated episode, and existing skill learning methods either focus on single-task repetition or involve complex, multi-stage pipelines for skill extraction, retrieval, and execution. This paper introduces SkillRise, a streamlined reinforcement learning framework designed for learning and evolving skills across multiple tasks. SkillRise organizes related task instances into progressively challenging sequences. A single policy within the framework dynamically alternates between solving the current task and curating an evolving "skill document," which is then directly passed to the subsequent task. The framework employs a decoupled credit assignment mechanism, where task solving is supervised by the immediate task outcome, and skill curation is guided by discounted downstream outcomes. Experiments conducted on ALFWorld, WebShop, and ScienceWorld benchmarks show that SkillRise outperforms comparative methods, achieving significant gains in Pass@1 performance. The research indicates that SkillRise effectively reuses transferable skills across tasks, demonstrating improved performance with longer sequences of related tasks and substantially reducing the runtime overhead associated with multi-stage skill learning pipelines.