SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning

arXiv cs.CL · 6 hours ago

Researchers introduced SEED, a framework that trains large language models as agents by extracting reusable skills from completed trajectories and using these skills to provide denser learning signals during reinforcement learning. The method fine-tunes policies to analyze trajectories and generate natural-language skills that capture workflows and failure-avoidance patterns, then re-scores actions to create token-level supervision signals. SEED demonstrated consistent improvements in performance and sample efficiency on text-based and vision-based tasks, with code released on GitHub.