{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-30-014",
    "source_story_id": "tmp-story-evo-wam-verified-rollouts",
    "edition_id": "mp-2026-09-30-morning-0083",
    "edition_url": "https://themachinepress.com/edition/2026-09-30",
    "position": 14,
    "story_type": "dispatch",
    "section": "robotics",
    "editorial_classification": "editorial",
    "headline": "A World-Action Model Learned From Its Own Verified Rollouts",
    "slug": "a-world-action-model-learned-from-its-own-verified-rollouts",
    "dek": "EVO-WAM kept generated prefixes only when video showed completion and inverse dynamics agreed with the actions.",
    "summary": "EVO-WAM kept generated prefixes only when video showed completion and inverse dynamics agreed with the actions.",
    "body_text": "EVO-WAM adapts a world-action model to unseen tasks without executing each candidate rollout externally. It selects prefixes that a vision-language model judges task-completing, checks video-action consistency with inverse dynamics and retrains iteratively on the verified subset. Across seven unseen RoboTwin 2.0 tasks, reported average success rose from 26.9% to 68.0% for Cosmos3 and from 28.5% to 46.4% for DreamZero. On three real long-horizon composites, Cosmos3 rose from 20.0% to 76.7%. The verifier stack can still inherit model error, so external validation remains necessary.",
    "why_it_matters": "EVO-WAM kept generated prefixes only when video showed completion and inverse dynamics agreed with the actions.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-30-014/a-world-action-model-learned-from-its-own-verified-rollouts",
    "json_url": "https://themachinepress.com/story/mp-2026-09-30-014.json",
    "first_published_at": "2026-09-30T09:00:00.000-04:00",
    "modified_at": "2026-09-30T09:02:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-30-014-001",
        "text": "EVO-WAM kept generated prefixes only when video showed completion and inverse dynamics agreed with the actions.",
        "source_ids": [
          "source-2026-09-30-014"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-30-014"
    ],
    "tags": [
      "world-action models",
      "self-improvement",
      "trajectory verification"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-30-014",
      "title": "arXiv preprint 2609.38057",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.38057",
      "canonical_url": "https://arxiv.org/abs/2609.38057",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-29T13:25:35.000-04:00",
      "accessed_at": "2026-09-30T08:24:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-30-014-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "A World-Action Model Learned From Its Own Verified Rollouts",
    "publisher": "The Machine Press",
    "published_at": "2026-09-30T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-30-014/a-world-action-model-learned-from-its-own-verified-rollouts"
  }
}