{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-11-03-009",
    "source_story_id": "tmp-story-muon-adamw-reversal",
    "edition_id": "mp-2026-11-03-morning-0116",
    "edition_url": "https://themachinepress.com/edition/2026-11-03",
    "position": 9,
    "story_type": "dispatch",
    "section": "frontier-models",
    "editorial_classification": "editorial",
    "headline": "More Pretraining Reverses the Optimizer Winner",
    "slug": "more-pretraining-reverses-the-optimizer-winner",
    "dek": "Muon minimizes pretraining loss faster, but AdamW overtakes it after fine-tuning once the pretraining data budget grows large enough.",
    "summary": "Muon minimizes pretraining loss faster, but AdamW overtakes it after fine-tuning once the pretraining data budget grows large enough.",
    "body_text": "Muon has gained attention as a faster pretraining optimizer, but lower pretraining loss does not guarantee the strongest fine-tuned model. Across the paper's language-model experiments, AdamW moves ahead after fine-tuning beyond a sufficiently large data budget, and that crossover arrives at a smaller token-to-parameter ratio as model size increases. The authors propose a margin-to-spread ratio—correct-logit margin relative to the dispersion of wrong logits—to explain the reversal. Muon leads on that measure at small budgets and AdamW at large ones, matching the downstream pattern; a linear associative-memory analysis reproduces the direction after an update. The crossover depends on the tested regimes and is not a universal optimizer ranking.",
    "why_it_matters": "Muon minimizes pretraining loss faster, but AdamW overtakes it after fine-tuning once the pretraining data budget grows large enough.",
    "limitations": [
      "Muon has gained attention as a faster pretraining optimizer, but lower pretraining loss does not guarantee the strongest fine-tuned model.",
      "The crossover depends on the tested regimes and is not a universal optimizer ranking."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-11-03-009/more-pretraining-reverses-the-optimizer-winner",
    "json_url": "https://themachinepress.com/story/mp-2026-11-03-009.json",
    "first_published_at": "2026-11-03T09:00:00.000-05:00",
    "modified_at": "2026-11-03T09:00:00.000-05:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-11-03-009-001",
        "text": "Muon minimizes pretraining loss faster, but AdamW overtakes it after fine-tuning once the pretraining data budget grows large enough.",
        "source_ids": [
          "source-2026-11-03-009"
        ],
        "qualification": "Muon has gained attention as a faster pretraining optimizer, but lower pretraining loss does not guarantee the strongest fine-tuned model."
      }
    ],
    "source_ids": [
      "source-2026-11-03-009"
    ],
    "tags": [
      "optimization",
      "pretraining",
      "fine-tuning",
      "language models"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-11-03-009",
      "title": "arXiv preprint 2611.03633, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2611.03633v1",
      "canonical_url": "https://arxiv.org/abs/2611.03633v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-11-02T13:56:43.000-05:00",
      "accessed_at": "2026-11-03T08:18:17.000-05:00",
      "supports_claim_ids": [
        "claim-mp-2026-11-03-009-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "More Pretraining Reverses the Optimizer Winner",
    "publisher": "The Machine Press",
    "published_at": "2026-11-03T09:00:00.000-05:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-11-03-009/more-pretraining-reverses-the-optimizer-winner"
  }
}