SkillRise: Agentic RL Framework for Cross-Task Skill Learning
Researchers have unveiled SkillRise, a comprehensive reinforcement learning framework designed for skill acquisition across various tasks. This innovative system arranges similar instances into increasingly difficult sequences and employs a singular policy to switch between task resolution and the development of a dynamic skill document for subsequent tasks. It utilizes a decoupled credit assignment mechanism that oversees task resolution based on current outcomes and skill curation using discounted future results. Trials conducted on ALFWorld, WebShop, and ScienceWorld demonstrate that SkillRise outperforms other methods, achieving the highest Pass@1 performance. This framework overcomes the challenges of conventional agentic reinforcement learning, which views tasks as separate episodes, and existing skill learning techniques that either emphasize repetitive task attempts or adopt multi-stage pipelines.
Key facts
- SkillRise is a unified reinforcement learning framework for learning skills across tasks.
- It organizes related instances into progressively challenging sequences.
- It uses a single policy to alternate between task solving and curating an evolving skill document.
- Decoupled credit assignment supervises solving with current task outcome and curation with discounted downstream outcomes.
- Experiments were conducted on ALFWorld, WebShop, and ScienceWorld.
- SkillRise achieved the strongest Pass@1 performance among compared methods.
- Standard agentic RL treats tasks as independent episodes.
- Existing skill learning approaches focus on repeated attempts of one task or use multi-stage pipelines.
Entities
—