{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-25-009",
    "source_story_id": "tmp-story-srpo-reflection-patches",
    "edition_id": "mp-2026-08-25-morning-0047",
    "edition_url": "https://themachinepress.com/edition/2026-08-25",
    "position": 9,
    "story_type": "dispatch",
    "section": "research",
    "editorial_classification": "editorial",
    "headline": "The Model Turned Its Own Failure Into the Training Signal",
    "slug": "the-model-turned-its-own-failure-into-the-training-signal",
    "dek": "SRPO converts completed trajectories into concise reflection patches and dense token-level supervision.",
    "summary": "SRPO converts completed trajectories into concise reflection patches and dense token-level supervision.",
    "body_text": "With a Qwen3-8B base, the authors report 73.3 percent on AIME 2024 using 8 percent of the training FLOPs of scaled supervised fine-tuning, alongside gains on WebShop, ALFWorld and SWE-Bench-Lite. The method uses reflection-conditioned teacher scores without a separate critic or larger teacher model.",
    "why_it_matters": "SRPO converts completed trajectories into concise reflection patches and dense token-level supervision.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-25-009/the-model-turned-its-own-failure-into-the-training-signal",
    "json_url": "https://themachinepress.com/story/mp-2026-08-25-009.json",
    "first_published_at": "2026-08-25T09:00:00.000-04:00",
    "modified_at": "2026-08-25T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-08-25-009-001",
        "text": "SRPO converts completed trajectories into concise reflection patches and dense token-level supervision.",
        "source_ids": [
          "source-2026-08-25-009"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-08-25-009"
    ],
    "tags": [
      "reinforcement learning",
      "self-reflection",
      "long-horizon reasoning"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-25-009",
      "title": "arXiv preprint 2608.23493",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.23493",
      "canonical_url": "https://arxiv.org/abs/2608.23493",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-08-24T12:55:09.000-04:00",
      "accessed_at": "2026-08-25T08:17:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-25-009-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "The Model Turned Its Own Failure Into the Training Signal",
    "publisher": "The Machine Press",
    "published_at": "2026-08-25T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-25-009/the-model-turned-its-own-failure-into-the-training-signal"
  }
}