{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-26-005",
    "source_story_id": "tmp-story-poem-predict-rl-policy",
    "edition_id": "mp-2026-09-26-morning-0079",
    "edition_url": "https://themachinepress.com/edition/2026-09-26",
    "position": 5,
    "story_type": "dispatch",
    "section": "frontier-models",
    "editorial_classification": "editorial",
    "headline": "A New Reward Didn’t Need a New RL Run",
    "slug": "a-new-reward-didn-t-need-a-new-rl-run",
    "dek": "PoEM approximated a target post-training policy from models already optimized on other rewards.",
    "summary": "PoEM approximated a target post-training policy from models already optimized on other rewards.",
    "body_text": "The authors show that when a new reward is a linear combination of known rewards, its reinforcement-learned policy can also be combined in log-policy space. They further observe an approximately low-rank structure even when rewards are not linearly related, then estimate weights from reward or basis-policy outputs rather than launching another full training run. Experiments span synthetic and real rewards in text and image settings. PoEM predicts the outcome of the studied optimization setups; it does not eliminate the need to validate an approximated policy before deployment.",
    "why_it_matters": "PoEM approximated a target post-training policy from models already optimized on other rewards.",
    "limitations": [
      "They further observe an approximately low-rank structure even when rewards are not linearly related, then estimate weights from reward or basis-policy outputs rather than launching another full training run.",
      "PoEM predicts the outcome of the studied optimization setups; it does not eliminate the need to validate an approximated policy before deployment."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-26-005/a-new-reward-didn-t-need-a-new-rl-run",
    "json_url": "https://themachinepress.com/story/mp-2026-09-26-005.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-2026-09-26-005-001",
        "text": "PoEM approximated a target post-training policy from models already optimized on other rewards.",
        "source_ids": [
          "source-2026-09-26-005"
        ],
        "qualification": "They further observe an approximately low-rank structure even when rewards are not linearly related, then estimate weights from reward or basis-policy outputs rather than launching another full training run."
      }
    ],
    "source_ids": [
      "source-2026-09-26-005"
    ],
    "tags": [
      "reinforcement learning",
      "post-training",
      "policy approximation"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-26-005",
      "title": "arXiv preprint 2609.30226",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.30226",
      "canonical_url": "https://arxiv.org/abs/2609.30226",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-24T13:50:25.000-04:00",
      "accessed_at": "2026-09-26T08:20:31.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-26-005-001"
      ]
    }
  ],
  "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 New Reward Didn’t Need a New RL Run",
    "publisher": "The Machine Press",
    "published_at": "2026-09-26T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-26-005/a-new-reward-didn-t-need-a-new-rl-run"
  }
}