{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-26-008",
    "source_story_id": "tmp-story-minimally-invasive-steering",
    "edition_id": "mp-2026-09-26-morning-0079",
    "edition_url": "https://themachinepress.com/edition/2026-09-26",
    "position": 8,
    "story_type": "dispatch",
    "section": "frontier-models",
    "editorial_classification": "editorial",
    "headline": "Steering Changed Less of the Language Model",
    "slug": "steering-changed-less-of-the-language-model",
    "dek": "MISVO led mean reward in six of seven model-task settings while keeping diversity and coherence near Best-of-N.",
    "summary": "MISVO led mean reward in six of seven model-task settings while keeping diversity and coherence near Best-of-N.",
    "body_text": "Pre-logit steering can raise a test-time reward but also distort the rest of a frozen model's output distribution. MISVO penalizes interventions using the local Fisher geometry of token probabilities and optimizes position-specific vectors without updating model weights. Across preference and code-generation tasks on roughly one- to fourteen-billion-parameter models, the authors report higher reward with restrained distributional change. The comparison is limited to the selected models, tasks and reward functions.",
    "why_it_matters": "MISVO led mean reward in six of seven model-task settings while keeping diversity and coherence near Best-of-N.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-26-008/steering-changed-less-of-the-language-model",
    "json_url": "https://themachinepress.com/story/mp-2026-09-26-008.json",
    "first_published_at": "2026-09-26T09:00:00.000-04:00",
    "modified_at": "2026-09-26T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-26-008-001",
        "text": "MISVO led mean reward in six of seven model-task settings while keeping diversity and coherence near Best-of-N.",
        "source_ids": [
          "source-2026-09-26-008"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-26-008"
    ],
    "tags": [
      "model steering",
      "Fisher geometry",
      "test-time adaptation"
    ],
    "image_url": "https://themachinepress.com/issues/2026-09-26/minimal-steering-wireframe.webp",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-26-008",
      "title": "arXiv preprint 2609.30218",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.30218",
      "canonical_url": "https://arxiv.org/abs/2609.30218",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-24T13:46:46.000-04:00",
      "accessed_at": "2026-09-26T08:20:31.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-26-008-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Steering Changed Less of the Language Model",
    "publisher": "The Machine Press",
    "published_at": "2026-09-26T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-26-008/steering-changed-less-of-the-language-model"
  }
}