SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution

2026-07-29

AI Summary

This work addresses the challenge that large language model agents struggle to efficiently extract, reuse, and evolve transferable skills across tasks, as existing reinforcement learning approaches typically treat tasks in isolation and employ multi-stage, tightly coupled skill-learning pipelines that suffer from low efficiency. To overcome these limitations, the authors propose SkillRise, a framework that constructs a progressive task sequence and employs a single policy to alternately solve tasks and dynamically curate skill documentation. Optimized skills are directly transferred to subsequent tasks, enabling end-to-end cross-task skill evolution. SkillRise unifies skill extraction, refinement, and reuse while decoupling cross-task credit assignment, yielding scalability that improves with the length of the task sequence. Experiments demonstrate that SkillRise achieves 2.3–8.5 percentage points higher Pass@1 performance than the strongest baselines on ALFWorld, WebShop, and ScienceWorld, significantly reduces training overhead, and enables effective skill reuse within a single attempt.

Abstract

Large language model agents often encounter related yet distinct tasks that share reusable solution patterns. Yet standard agentic reinforcement learning treats tasks as independent episodes, while existing approaches to skill learning either focus on repeated attempts of one task or use pipelines with multiple stages that entangle extraction, retrieval, and execution. We introduce SkillRise, a unified reinforcement learning framework for learning skills across tasks. SkillRise organizes related instances into progressively challenging sequences and uses a single policy to alternate between task solving and curating an evolving skill document passed directly to the next task. Decoupled credit assignment across tasks supervises solving with the current task outcome and curation with discounted downstream outcomes. Experiments on ALFWorld, WebShop, and ScienceWorld show that SkillRise achieves the strongest Pass@1 performance among the compared methods, with gains over the strongest baseline ranging from 2.3 to 8.5 percentage points. Although trained across distinct tasks, its learned curation policy remains effective for repeated attempts on the same task. Further analysis reveals scaling at test time across tasks: performance improves with longer sequences of related tasks even when each task is attempted only once. This trend suggests that SkillRise reuses transferable skills across tasks rather than benefiting from repeated sampling of the same task. SkillRise further retains strong performance while substantially reducing the runtime overhead of skill learning pipelines with multiple stages. Together, these results provide a simple and efficient training paradigm for LLM agents to extract, refine, and reuse transferable skills across tasks.

Problem

• cross-task skill learning
• agentic reinforcement learning
• skill reuse
• transferable skills
• large language model agents

Innovation

• agentic reinforcement learning
• cross-task skill transfer
• skill curation
• decoupled credit assignment
• LLM agents