{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-20-019",
    "source_story_id": "tmp-brief-tokengate-prefill",
    "edition_id": "mp-2026-10-20-morning-0102",
    "edition_url": "https://themachinepress.com/edition/2026-10-20",
    "position": 21,
    "story_type": "ticker",
    "section": "infrastructure",
    "editorial_classification": "editorial",
    "headline": "The Model Skims Before It Reads",
    "slug": "the-model-skims-before-it-reads",
    "dek": "TokenGate learns which long-context tokens can stop moving through later layers during prefill.",
    "summary": "TokenGate learns which long-context tokens can stop moving through later layers during prefill.",
    "body_text": "TokenGate adds a lightweight importance module and token-gated attention so existing models can prune input during prefill. Across reported tasks, time to first token improved by 1.4 to 11.3 times and peak memory fell by 1.5 to 3.4 times without measured accuracy loss. Gains depend on model, task and hardware, but the component extends both standard and hybrid-attention transformers through parameter-efficient adaptation.",
    "why_it_matters": "TokenGate learns which long-context tokens can stop moving through later layers during prefill.",
    "limitations": [],
    "importance": 7,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-20-019/the-model-skims-before-it-reads",
    "json_url": "https://themachinepress.com/story/mp-2026-10-20-019.json",
    "first_published_at": "2026-10-20T09:00:00.000-04:00",
    "modified_at": "2026-10-20T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-20-019-001",
        "text": "TokenGate learns which long-context tokens can stop moving through later layers during prefill.",
        "source_ids": [
          "source-2026-10-20-021"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-10-20-021"
    ],
    "tags": [
      "inference",
      "token pruning",
      "memory efficiency"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-20-021",
      "title": "arXiv preprint 2610.26754, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.26754v1",
      "canonical_url": "https://arxiv.org/abs/2610.26754v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-19T13:15:53.000-04:00",
      "accessed_at": "2026-10-20T08:15:06.176-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-20-019-001"
      ]
    }
  ],
  "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 Model Skims Before It Reads",
    "publisher": "The Machine Press",
    "published_at": "2026-10-20T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-20-019/the-model-skims-before-it-reads"
  }
}