{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-11-01-001",
    "source_story_id": "tmp-lead-data-attribution",
    "edition_id": "mp-2026-11-01-morning-0114",
    "edition_url": "https://themachinepress.com/edition/2026-11-01",
    "position": 1,
    "story_type": "lead",
    "section": "benchmarks-evals",
    "editorial_classification": "editorial",
    "headline": "The Evidence Was There. Averaging Hid It.",
    "slug": "the-evidence-was-there-averaging-hid-it",
    "dek": "A controlled benchmark found that causal training evidence was highly localized—and that changing the gradient and aggregation choices raised ground-truth recall by 35 to 72 percentage points.",
    "summary": "A controlled benchmark found that causal training evidence was highly localized—and that changing the gradient and aggregation choices raised ground-truth recall by 35 to 72 percentage points.",
    "body_text": "Language-model data attribution asks which training records caused a particular output, but ordinary evaluations often change several properties at once. CATT instead fine-tunes models on fictitious facts with known provenance, then varies redundancy and lexical distractors separately. Both conditions degraded attribution. The researchers found that causal evidence within a document clustered in a few tokens, while whole-document averaging diluted it; negative-log-odds gradients plus top-token aggregation improved ground-truth recall by 35 to 72 percentage points across tested baselines. On FTRACE-TREX, the gradient change improved counterfactual-retraining performance by as much as 4.9 times. Those gains diagnose the tested methods and datasets, not a guarantee that any production model's training history can be reconstructed.",
    "why_it_matters": "A controlled benchmark found that causal training evidence was highly localized—and that changing the gradient and aggregation choices raised ground-truth recall by 35 to 72 percentage points.",
    "limitations": [
      "Those gains diagnose the tested methods and datasets, not a guarantee that any production model's training history can be reconstructed."
    ],
    "importance": 10,
    "canonical_url": "https://themachinepress.com/story/mp-2026-11-01-001/the-evidence-was-there-averaging-hid-it",
    "json_url": "https://themachinepress.com/story/mp-2026-11-01-001.json",
    "first_published_at": "2026-11-01T09:00:00.000-05:00",
    "modified_at": "2026-11-01T09:00:00.000-05:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-11-01-001-001",
        "text": "A controlled benchmark found that causal training evidence was highly localized—and that changing the gradient and aggregation choices raised ground-truth recall by 35 to 72 percentage points.",
        "source_ids": [
          "source-2026-11-01-001"
        ],
        "qualification": "Those gains diagnose the tested methods and datasets, not a guarantee that any production model's training history can be reconstructed."
      }
    ],
    "source_ids": [
      "source-2026-11-01-001"
    ],
    "tags": [
      "data attribution",
      "language models",
      "evaluation",
      "training data"
    ],
    "image_url": "https://themachinepress.com/issues/2026-11-01/lead-data-attribution.png",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-11-01-001",
      "title": "arXiv preprint 2610.40655, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.40655v1",
      "canonical_url": "https://arxiv.org/abs/2610.40655v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-28T20:00:00.000-04:00",
      "accessed_at": "2026-11-01T08:20:00.000-05:00",
      "supports_claim_ids": [
        "claim-mp-2026-11-01-001-001"
      ]
    }
  ],
  "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 Evidence Was There. Averaging Hid It.",
    "publisher": "The Machine Press",
    "published_at": "2026-11-01T09:00:00.000-05:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-11-01-001/the-evidence-was-there-averaging-hid-it"
  }
}