{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-12-027",
    "source_story_id": "tmp-story-learned-answer-selection-abstention",
    "edition_id": "mp-2026-10-12-morning-0094",
    "edition_url": "https://themachinepress.com/edition/2026-10-12",
    "position": 16,
    "story_type": "dispatch",
    "section": "benchmarks-evals",
    "editorial_classification": "editorial",
    "headline": "The Model Learned to Withhold Some Answers. The Guarantee Did Not Travel.",
    "slug": "the-model-learned-to-withhold-some-answers-the-guarantee-did-not-travel",
    "dek": "A separate answer selector improved one factual dataset but still lost to silence on another.",
    "summary": "A separate answer selector improved one factual dataset but still lost to silence on another.",
    "body_text": "Researchers asked whether reliable abstention should come from training the language model itself or from a separate component that checks its proposed answers. On Qwen3-14B, the learned selector improved threshold-averaged utility over the original model’s calibrated confidence filter. Under exact-match grading on TriviaQA, mean coverage rose from 20.2% to 30.9%, while the pooled error rate among scored answers fell from 23.0% to 14.5%. The tested reinforcement-learning changes did not produce consistent gains. On NQ-Open, the selector improved on the filter but still performed worse than always abstaining. A partial human-reviewed audit preserved the TriviaQA finding. The contrast matters for deploying a system that must decide whether it knows enough: selecting answers can help in a particular setting, but one dataset’s improvement does not establish reliable uncertainty handling elsewhere. The authors leave that broader problem unresolved.",
    "why_it_matters": "A separate answer selector improved one factual dataset but still lost to silence on another.",
    "limitations": [
      "The tested reinforcement-learning changes did not produce consistent gains.",
      "The contrast matters for deploying a system that must decide whether it knows enough: selecting answers can help in a particular setting, but one dataset’s improvement does not establish reliable uncertainty handling elsewhere."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-12-027/the-model-learned-to-withhold-some-answers-the-guarantee-did-not-travel",
    "json_url": "https://themachinepress.com/story/mp-2026-10-12-027.json",
    "first_published_at": "2026-10-12T09:00:00.000-04:00",
    "modified_at": "2026-10-12T09:48:10.481-04:00",
    "content_status": "updated",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-12-027-001",
        "text": "A separate answer selector improved one factual dataset but still lost to silence on another.",
        "source_ids": [
          "source-2026-10-12-016"
        ],
        "qualification": "The tested reinforcement-learning changes did not produce consistent gains."
      }
    ],
    "source_ids": [
      "source-2026-10-12-016"
    ],
    "tags": [
      "abstention",
      "answer-selection",
      "calibration",
      "preprint"
    ],
    "image_url": null,
    "corrections": [
      {
        "correction_id": "correction-2026-10-12-001",
        "revision": 2,
        "summary": "Publication note: The October 12, 2026 edition was prepared and published late on October 12 after the scheduled newsroom task could not start. The edition retains its October 12 date and 9:00 AM America/Detroit reveal time. Reports use sources available before that cutoff; actual preparation and image-generation times are preserved in the provenance records.",
        "published_at": "2026-10-12T09:48:10.481-04:00",
        "affected_story_ids": [],
        "affected_claim_ids": []
      }
    ]
  },
  "sources": [
    {
      "source_id": "source-2026-10-12-016",
      "title": "arXiv preprint 2610.13821, original v1",
      "publisher": "arXiv / submitting research authors",
      "url": "https://arxiv.org/abs/2610.13821v1",
      "canonical_url": "https://arxiv.org/abs/2610.13821v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-09T12:33:27.000-04:00",
      "accessed_at": "2026-10-12T09:32:36.812-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-12-027-001"
      ]
    }
  ],
  "corrections": [
    {
      "correction_id": "correction-2026-10-12-001",
      "revision": 2,
      "summary": "Publication note: The October 12, 2026 edition was prepared and published late on October 12 after the scheduled newsroom task could not start. The edition retains its October 12 date and 9:00 AM America/Detroit reveal time. Reports use sources available before that cutoff; actual preparation and image-generation times are preserved in the provenance records.",
      "published_at": "2026-10-12T09:48:10.481-04:00",
      "affected_story_ids": [],
      "affected_claim_ids": []
    }
  ],
  "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 Learned to Withhold Some Answers. The Guarantee Did Not Travel.",
    "publisher": "The Machine Press",
    "published_at": "2026-10-12T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-12-027/the-model-learned-to-withhold-some-answers-the-guarantee-did-not-travel"
  }
}