{
  "id": 3668,
  "url": "https://arxiv.org/abs/2605.27140v1",
  "title": "StepOPSD: Step-Aware Online Preference Distillation for Agent Reinforcement Learning",
  "summary": "Reinforcement learning for multi-turn agents suffers from a credit-assignment mismatch: rewards are sparse and trajectory-level, while success often hinges on a few local decisions. Existing online policy distillation (OPD) provides denser token-level supervision, but typically treats heterogeneous agent trajectories as monolithic strings rather than causal interaction units. We present StepOPSD, a post-rollout preference self-distillation framework that takes the agent step as the unit of credi",
  "authors": "Yanfei Zhang, Xu Lin, Chenglin Wu",
  "category": "research",
  "topics": "regulation,agents-autonomy",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-26T15:07:03.000Z",
  "fetched_at": "2026-07-14T16:30:27.607Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/3668",
  "original_url": "https://arxiv.org/abs/2605.27140v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}