AIAny
Icon for item

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

Converts completed on-policy trajectories into natural-language 'hindsight skills' and converts the skill-induced action probability shifts into a dense token-level on-policy distillation signal, jointly optimized with outcome-based RL to improve sample efficiency and long-horizon agent behavior.

Introduction

Most outcome-based RL for interactive agents trains from sparse episode returns, leaving little guidance for intermediate token-level decisions. SEED flips that gap into a source of supervision: the policy itself analyzes finished trajectories to extract reusable, natural-language skills, then distills the behavioral effect of those skills back into the policy as a dense on-policy supervisory signal that co-evolves with the policy distribution.

Key Findings
  • Hindsight-skill extraction: Treating completed trajectories as analyzable artifacts yields compact natural-language skills that capture reusable workflows, decisive observations, or failure-avoidance heuristics — these are used as alternate contexts to re-score actions.
  • On-policy distillation signal: The method converts the probability shift induced by skill-augmented contexts into a token-level distillation target, producing dense supervision aligned with the sampled trajectory distribution instead of relying only on sparse episode rewards.
  • Self-evolving loop: Because the current policy both collects trajectories and plays the role of analyzer, skill generation and policy improvement bootstrap each other; as the policy improves, extracted skills become more relevant to new data.
  • Empirical benefits: Across text-based and vision-based agentic tasks, the approach improves sample efficiency and generalization to unseen scenarios compared with outcome-only RL baselines.
Who it fits and trade-offs

Great fit if you train interactive LLM agents or multimodal agents on long-horizon tasks where trajectory-level rewards are sparse and you can afford additional compute for trajectory analysis and distillation. It is less suitable if you cannot reliably collect complete on-policy trajectories, if running extra analysis per trajectory is prohibitively expensive, or if you require pure offline policy distillation without online data collection.

How it works (brief)

SEED fine-tunes the policy to act as an analyzer that, given a finished trajectory, writes natural-language skills that summarize actionable patterns. During RL rollout the same analyzer generates skills from recent trajectories; the system re-scores sampled actions under ordinary vs skill-augmented contexts and converts the induced logit/probability differences into a dense token-level loss term that is optimized jointly with the outcome-based RL objective. The pipeline keeps the auxiliary supervision synchronized with the current trajectory distribution rather than using static hindsight heuristics.

Information

  • Websitearxiv.org
  • AuthorsJinyang Wu, Shuo Yang, Zhengxi Lu, Fan Zhang, Yuhao Shen, Lang Feng, Haoran Luo, Zheng Lian, Shuai Zhang, Zhengqi Wen
  • Published date2026/07/16

More Items

Analyzes adversarial weaknesses of World-Action Models (WAMs) via BadWAM, a framework that crafts visual perturbations to decouple a model’s imagined future from its executed actions. Introduces two attack modes—action-only (disruptive) and imagination-preserving (stealthy)—and shows large drops in closed-loop task success (e.g., 96.5%→43.1%).

Turns fragile, implicit search progress into explicit, persistent, shared state for multi-agent information seeking — externalizes progress as Frontier Task, Evidence Graph, Coverage Map and Failure Memory, and uses pipeline-parallel scheduling plus a middleware harness to avoid repeated failed searches and improve utilization and throughput.

Acquires repository knowledge via a targeted QA loop before generating patches, decoupling knowledge acquisition from repair. A Questioner and Answerer produce evidence-grounded QA pairs that a Resolver uses to generate fixes; improves Pass@1 on SWE-bench Verified with modest overhead.