{
  "schema_version": "1.1",
  "id": "s13:https://arxiv.org/abs/2609.04172v1",
  "slug": "2609-04172v1-0kofdqa",
  "url": "https://feed7.dev/p/2609-04172v1-0kofdqa",
  "title": "Rethinking On-Policy Distillation of Large Language Models II: One Training Example",
  "why_included": "On-policy distillation recovered most full-data gains from one query; diverse rollouts mattered more than dataset size, while slow student alignment remained the bottleneck.",
  "summary": "A single-query OPD run reached **71.5% state coverage** and recovered most of full-data OPD’s gain across tested tasks and model families. With **16 queries**, coverage rose to **98.9%** and matched full-data training.",
  "practical_implication": "If you distill smaller models for agent workloads, optimize prompts for diverse visited states before collecting a large task dataset. The experiments suggest rollout coverage, not query count or topical content alone, is the useful data signal.",
  "agent_context": "A single-query OPD run reached **71.5% state coverage** and recovered most of full-data OPD’s gain across tested tasks and model families. With **16 queries**, coverage rose to **98.9%** and matched full-data training.\n\nIf you distill smaller models for agent workloads, optimize prompts for diverse visited states before collecting a large task dataset. The experiments suggest rollout coverage, not query count or topical content alone, is the useful data signal.\n\nThis does not make training cheap: student-teacher alignment still took **hundreds of steps**, even on fixed states. The claims concern OPD experiments, so deployment quality and transfer to a specific coding-agent workload still need validation.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2609.04172v1",
    "published_at": "2026-09-03T17:54:38.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "model",
  "domains": [],
  "topics": [
    "reasoning",
    "model-selection"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "This does not make training cheap: student-teacher alignment still took **hundreds of steps**, even on fixed states. The claims concern OPD experiments, so deployment quality and transfer to a specific coding-agent workload still need validation."
  ],
  "connected_context": {
    "meaning": "This sharply narrows data-efficiency claims for distillation: topical dataset breadth may be unnecessary when a few prompts generate broad state coverage, but optimization remains costly because alignment still requires hundreds of steps. For agent adaptation, prompt selection should be judged by visited-state diversity, not query count alone.",
    "corpus_size": 691,
    "generated_at": "2026-09-05T10:08:08.498Z",
    "connections": [
      {
        "title": "Data Quality Is the Compute Multiplier — Ari Morcos, DatologyAI",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=_PdK6x7PQNM",
        "feed7_url": "https://feed7.dev/p/data-quality-is-the-compute-multiplier-ari-morcos-datologyai-0x7k2ve",
        "reason": "Makes the general data-curation principle operational for OPD by identifying rollout-state coverage as the relevant information-value signal."
      },
      {
        "title": "OctoLong: Mid-Training On Cross-Repository Code Contexts Enhances Long-Context Modeling",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.05141v1",
        "feed7_url": "https://feed7.dev/p/2608-05141v1-0kai3a6",
        "reason": "Both favor structurally informative training exposure over raw volume: dependency-linked contexts in mid-training and diverse visited states in distillation."
      },
      {
        "title": "Consolidating RLVR Capabilities Across Domains: A Deep Dive into Fusion Paradigms",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.27409v1",
        "feed7_url": "https://feed7.dev/p/2608-27409v1-1bivffd",
        "reason": "Adds a useful metric contrast: OPD gains track visited-state coverage, while RLVR fusion did not broaden solution coverage despite similar average performance."
      },
      {
        "title": "Adaption Labs: Gradient-Free Continual Learning — Sara Hooker, Adaption",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=XEd_SRVHBgU",
        "feed7_url": "https://feed7.dev/p/adaption-labs-gradient-free-continual-learning-sara-hooker-adaption-1pozjjc",
        "reason": "Supplies a concrete constraint for automated adaptation systems: discovering a tiny effective prompt set would not remove the substantial student-alignment compute."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-09-03T17:54:38.000Z",
  "modified_at": "2026-09-03T17:54:38.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2609-04172v1-0kofdqa",
    "json": "https://feed7.dev/p/2609-04172v1-0kofdqa.json",
    "markdown": "https://feed7.dev/p/2609-04172v1-0kofdqa.md"
  }
}