{
  "id": 3676,
  "url": "https://arxiv.org/abs/2605.27028v1",
  "title": "Less is More: Early Stopping Rollout for On-Policy Distillation",
  "summary": "On-policy distillation has recently emerged as a promising alternative to standard sequence-level imitation, training a student by scoring its own rollouts with a teacher model. However, we observe ``Off-policy Teacher Decay'' problem in this paradigm: for the later tokens, with student's earlier trajectory as context that is off-policy to the teacher, the teacher's ability to produce a corrective score would decay, and may fall back to token-completion behavior learned in the pre-training stage",
  "authors": "Zhou Ziheng, Jiaqi Li, Huacong Tang, Ying Nian Wu, Demetri Terzopoulos",
  "category": "research",
  "topics": "regulation,children-education",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-26T13:49:37.000Z",
  "fetched_at": "2026-07-14T16:30:27.608Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/3676",
  "original_url": "https://arxiv.org/abs/2605.27028v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}