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

SEED tackles the sparse-reward problem in outcome-based RL for agentic LLMs by having the policy extract reusable natural-language 'skills' (workflows and failure-avoidance patterns) from its own completed trajectories, then using the action-probability gap between skill-augmented and unaugmented rollouts as a dense, token-level on-policy distillation signal. Reports gains on both text and vision agentic tasks plus better generalization to unseen scenarios.