SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
SkillRL introduces a framework to improve large language model agents by learning and evolving skills from experience rather than just storing raw data. It uses an experience-based distillation process to create a compact, hierarchical skill library (SKILLBANK) and a recursive evolution mechanism where the agent's policy and the skill library grow together during reinforcement learning. This results in agents that use structured knowledge efficiently, significantly reducing context size while boosting performance on complex tasks.