{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-25-006",
    "source_story_id": "tmp-story-safety-direction-penalty",
    "edition_id": "mp-2026-08-25-morning-0047",
    "edition_url": "https://themachinepress.com/edition/2026-08-25",
    "position": 6,
    "story_type": "dispatch",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "Reasoning Training Moved Along the Safety Direction",
    "slug": "reasoning-training-moved-along-the-safety-direction",
    "dek": "A representation-space audit links some reasoning fine-tunes to safety shifts and proposes a penalty at the implicated layers.",
    "summary": "A representation-space audit links some reasoning fine-tunes to safety shifts and proposes a penalty at the implicated layers.",
    "body_text": "The authors stress that reasoning-induced misalignment does not appear across every architecture, scale or dataset. On Qwen2.5 3B and 7B experiments, a learned safety-direction penalty restored measured safety while preserving benchmark reasoning performance, with diagnostics guiding which layers to include.",
    "why_it_matters": "A representation-space audit links some reasoning fine-tunes to safety shifts and proposes a penalty at the implicated layers.",
    "limitations": [
      "The authors stress that reasoning-induced misalignment does not appear across every architecture, scale or dataset."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-25-006/reasoning-training-moved-along-the-safety-direction",
    "json_url": "https://themachinepress.com/story/mp-2026-08-25-006.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-2026-08-25-006-001",
        "text": "A representation-space audit links some reasoning fine-tunes to safety shifts and proposes a penalty at the implicated layers.",
        "source_ids": [
          "source-2026-08-25-006"
        ],
        "qualification": "The authors stress that reasoning-induced misalignment does not appear across every architecture, scale or dataset."
      }
    ],
    "source_ids": [
      "source-2026-08-25-006"
    ],
    "tags": [
      "reasoning models",
      "safety",
      "representation geometry"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-25-006",
      "title": "arXiv preprint 2608.23497",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.23497",
      "canonical_url": "https://arxiv.org/abs/2608.23497",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-08-24T12:57:28.000-04:00",
      "accessed_at": "2026-08-25T08:17:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-25-006-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Reasoning Training Moved Along the Safety Direction",
    "publisher": "The Machine Press",
    "published_at": "2026-08-25T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-25-006/reasoning-training-moved-along-the-safety-direction"
  }
}