{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-06-012",
    "source_story_id": "tmp-story-one-example-on-policy-distillation",
    "edition_id": "mp-2026-09-06-morning-0059",
    "edition_url": "https://themachinepress.com/edition/2026-09-06",
    "position": 12,
    "story_type": "dispatch",
    "section": "frontier-models",
    "editorial_classification": "editorial",
    "headline": "One Training Query Reached Seventy-One Percent of the Teacher's States",
    "slug": "one-training-query-reached-seventy-one-percent-of-the-teacher-s-states",
    "dek": "On-policy distillation kept improving for hundreds of steps even when the student repeatedly learned from a single prompt.",
    "summary": "On-policy distillation kept improving for hundreds of steps even when the student repeatedly learned from a single prompt.",
    "body_text": "A single query's rollouts visited 71.5 percent of the states reached by full-data training, mostly within the first 100 steps. Sixteen semantically diverse queries reached 98.9 percent coverage and matched full-data gains, while content-light and off-domain prompts approached the real-query baseline. The authors argue that on-policy distillation is data-overfed but algorithm-starved: rollouts expose broad supervision quickly, then alignment absorbs it slowly. The finding is bounded to the tested tasks, teachers and model families.",
    "why_it_matters": "On-policy distillation kept improving for hundreds of steps even when the student repeatedly learned from a single prompt.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-06-012/one-training-query-reached-seventy-one-percent-of-the-teacher-s-states",
    "json_url": "https://themachinepress.com/story/mp-2026-09-06-012.json",
    "first_published_at": "2026-09-06T09:00:00.000-04:00",
    "modified_at": "2026-09-06T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-06-012-001",
        "text": "On-policy distillation kept improving for hundreds of steps even when the student repeatedly learned from a single prompt.",
        "source_ids": [
          "source-2026-09-06-012"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-06-012"
    ],
    "tags": [
      "distillation",
      "state coverage",
      "post-training"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-06-012",
      "title": "arXiv preprint 2609.04172",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.04172",
      "canonical_url": "https://arxiv.org/abs/2609.04172",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-03T13:54:38.000-04:00",
      "accessed_at": "2026-09-06T08:20:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-06-012-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "One Training Query Reached Seventy-One Percent of the Teacher's States",
    "publisher": "The Machine Press",
    "published_at": "2026-09-06T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-06-012/one-training-query-reached-seventy-one-percent-of-the-teacher-s-states"
  }
}