{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-29-004",
    "source_story_id": "tmp-story-telescopic-language-models",
    "edition_id": "mp-2026-09-29-morning-0082",
    "edition_url": "https://themachinepress.com/edition/2026-09-29",
    "position": 4,
    "story_type": "dispatch",
    "section": "frontier-models",
    "editorial_classification": "editorial",
    "headline": "One Language Model Ran at Every One of Its Twenty Depths",
    "slug": "one-language-model-ran-at-every-one-of-its-twenty-depths",
    "dek": "Stochastic prefix supervision trained a continuum of capacity instead of a few fixed exits.",
    "summary": "Stochastic prefix supervision trained a continuum of capacity instead of a few fixed exits.",
    "body_text": "Telescopic Language Models train one randomly truncated layer prefix alongside one full-capacity pass on every step. On a 200-million-parameter proxy trained over 20 billion FineWeb-Edu tokens, the resulting model remained usable at all 20 layer prefixes. The authors report a 43% to 44% reduction in area under the quality-budget curve versus fixed-exit suites, matching full-capacity quality at about 12% lower GPU cost per run. These are proxy-scale results, and deployment savings will depend on serving hardware and workload.",
    "why_it_matters": "Stochastic prefix supervision trained a continuum of capacity instead of a few fixed exits.",
    "limitations": [],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-29-004/one-language-model-ran-at-every-one-of-its-twenty-depths",
    "json_url": "https://themachinepress.com/story/mp-2026-09-29-004.json",
    "first_published_at": "2026-09-29T09:00:00.000-04:00",
    "modified_at": "2026-09-29T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-29-004-001",
        "text": "Stochastic prefix supervision trained a continuum of capacity instead of a few fixed exits.",
        "source_ids": [
          "source-2026-09-29-004"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-09-29-004"
    ],
    "tags": [
      "elastic inference",
      "language models",
      "compute budgets"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-29-004",
      "title": "arXiv preprint 2609.35769",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.35769",
      "canonical_url": "https://arxiv.org/abs/2609.35769",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-28T13:59:53.000-04:00",
      "accessed_at": "2026-09-29T08:18:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-29-004-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "One Language Model Ran at Every One of Its Twenty Depths",
    "publisher": "The Machine Press",
    "published_at": "2026-09-29T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-29-004/one-language-model-ran-at-every-one-of-its-twenty-depths"
  }
}