{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-08-018",
    "source_story_id": "tmp-sidebar-robormbench-paraphrase-rewards",
    "edition_id": "mp-2026-09-08-morning-0061",
    "edition_url": "https://themachinepress.com/edition/2026-09-08",
    "position": 20,
    "story_type": "ticker",
    "section": "safety",
    "editorial_classification": "editorial",
    "headline": "The Same Robot Motion Got Opposite Rewards After a Paraphrase",
    "slug": "the-same-robot-motion-got-opposite-rewards-after-a-paraphrase",
    "dek": "Across 2,390 trajectories and 21,673 verified rewrites, semantically equivalent goals could flip predicted progress from failure to success.",
    "summary": "Across 2,390 trajectories and 21,673 verified rewrites, semantically equivalent goals could flip predicted progress from failure to success.",
    "body_text": "ROBORMBENCH changes only the wording of a goal while holding the real-robot trajectory fixed. Reward instability appeared across proprietary and open models, grew with more divergent rewrites and was not reliably cured by scale or explicit reasoning; models trained with trajectory-grounded supervision were more stable.",
    "why_it_matters": "Across 2,390 trajectories and 21,673 verified rewrites, semantically equivalent goals could flip predicted progress from failure to success.",
    "limitations": [
      "ROBORMBENCH changes only the wording of a goal while holding the real-robot trajectory fixed.",
      "Reward instability appeared across proprietary and open models, grew with more divergent rewrites and was not reliably cured by scale or explicit reasoning; models trained with trajectory-grounded supervision were more stable."
    ],
    "importance": 7,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-08-018/the-same-robot-motion-got-opposite-rewards-after-a-paraphrase",
    "json_url": "https://themachinepress.com/story/mp-2026-09-08-018.json",
    "first_published_at": "2026-09-08T09:00:00.000-04:00",
    "modified_at": "2026-09-08T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-08-018-001",
        "text": "Across 2,390 trajectories and 21,673 verified rewrites, semantically equivalent goals could flip predicted progress from failure to success.",
        "source_ids": [
          "source-2026-09-08-020"
        ],
        "qualification": "ROBORMBENCH changes only the wording of a goal while holding the real-robot trajectory fixed."
      }
    ],
    "source_ids": [
      "source-2026-09-08-020"
    ],
    "tags": [
      "vision-language rewards",
      "paraphrase robustness",
      "robot learning"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-08-020",
      "title": "arXiv preprint 2609.05401",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.05401",
      "canonical_url": "https://arxiv.org/abs/2609.05401",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-04T13:47:58.000-04:00",
      "accessed_at": "2026-09-08T08:20:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-08-018-001"
      ]
    }
  ],
  "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 Same Robot Motion Got Opposite Rewards After a Paraphrase",
    "publisher": "The Machine Press",
    "published_at": "2026-09-08T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-08-018/the-same-robot-motion-got-opposite-rewards-after-a-paraphrase"
  }
}