{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-14-009",
    "source_story_id": "tmp-story-visual-token-budgeting",
    "edition_id": "mp-2026-10-14-morning-0096",
    "edition_url": "https://themachinepress.com/edition/2026-10-14",
    "position": 9,
    "story_type": "dispatch",
    "section": "chips-infrastructure",
    "editorial_classification": "editorial",
    "headline": "Nine Times Fewer Visual Tokens Preserve Most of the Score",
    "slug": "nine-times-fewer-visual-tokens-preserve-most-of-the-score",
    "dek": "A 1.6-million-parameter module assigns each image a query-aware token budget without retraining the base multimodal model.",
    "summary": "A 1.6-million-parameter module assigns each image a query-aware token budget without retraining the base multimodal model.",
    "body_text": "Multi-image questions can drown a multimodal model in visual tokens even when only a few regions matter. The mmVB module allocates a query-conditioned budget across images using cross-image context and does not require access to attention matrices at inference. The authors report ninefold fewer visual input tokens, fivefold faster time to first token and sevenfold fewer prefill operations while retaining 94.4 percent of full-resolution accuracy across eight benchmarks. Those figures are averages over the tested models and tasks, not a universal compression guarantee. The result suggests that multi-image systems should decide how much of each image to carry forward rather than compressing every frame equally.",
    "why_it_matters": "A 1.6-million-parameter module assigns each image a query-aware token budget without retraining the base multimodal model.",
    "limitations": [
      "Multi-image questions can drown a multimodal model in visual tokens even when only a few regions matter.",
      "The mmVB module allocates a query-conditioned budget across images using cross-image context and does not require access to attention matrices at inference.",
      "Those figures are averages over the tested models and tasks, not a universal compression guarantee."
    ],
    "importance": 7,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-14-009/nine-times-fewer-visual-tokens-preserve-most-of-the-score",
    "json_url": "https://themachinepress.com/story/mp-2026-10-14-009.json",
    "first_published_at": "2026-10-14T09:00:00.000-04:00",
    "modified_at": "2026-10-14T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-14-009-001",
        "text": "A 1.6-million-parameter module assigns each image a query-aware token budget without retraining the base multimodal model.",
        "source_ids": [
          "source-2026-10-14-009"
        ],
        "qualification": "Multi-image questions can drown a multimodal model in visual tokens even when only a few regions matter."
      }
    ],
    "source_ids": [
      "source-2026-10-14-009"
    ],
    "tags": [
      "mmVB",
      "multimodal models",
      "visual tokens"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-14-009",
      "title": "arXiv preprint 2610.18858, original v1",
      "publisher": "arXiv / submitting research authors",
      "url": "https://arxiv.org/abs/2610.18858v1",
      "canonical_url": "https://arxiv.org/abs/2610.18858v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-13T13:59:24.000-04:00",
      "accessed_at": "2026-10-14T08:24:30.364-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-14-009-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Nine Times Fewer Visual Tokens Preserve Most of the Score",
    "publisher": "The Machine Press",
    "published_at": "2026-10-14T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-14-009/nine-times-fewer-visual-tokens-preserve-most-of-the-score"
  }
}