{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-09-004",
    "source_story_id": "tmp-story-revit-depth-programmed-experts",
    "edition_id": "mp-2026-10-09-morning-0091",
    "edition_url": "https://themachinepress.com/edition/2026-10-09",
    "position": 4,
    "story_type": "dispatch",
    "section": "frontier-models",
    "editorial_classification": "editorial",
    "headline": "One Vision Block Reused Itself at Different Depths",
    "slug": "one-vision-block-reused-itself-at-different-depths",
    "dek": "reViT matched a full-depth encoder with about 70% fewer stored parameters, while keeping inference computation comparable.",
    "summary": "reViT matched a full-depth encoder with about 70% fewer stored parameters, while keeping inference computation comparable.",
    "body_text": "Reusing a Transformer block saves weights but can erase the transformations associated with different depths. reViT restores that variation by blending a small shared bank of feed-forward experts according to a continuous depth coordinate. Trained from scratch, reViT-B/16 matched DeiT III accuracy with about 70% fewer stored parameters at comparable inference FLOPs. An eight-expert model distilled from DINOv2 output features retained nearly all of its teacher’s linear-probe accuracy and transferred to segmentation and depth prediction. One checkpoint can run at several tested depths. Materializing it as a conventional dense graph removes online merging but expands storage. The reported saving is parameter storage, not a 70% reduction in inference work.",
    "why_it_matters": "reViT matched a full-depth encoder with about 70% fewer stored parameters, while keeping inference computation comparable.",
    "limitations": [
      "reViT restores that variation by blending a small shared bank of feed-forward experts according to a continuous depth coordinate.",
      "The reported saving is parameter storage, not a 70% reduction in inference work."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-09-004/one-vision-block-reused-itself-at-different-depths",
    "json_url": "https://themachinepress.com/story/mp-2026-10-09-004.json",
    "first_published_at": "2026-10-09T09:00:00.000-04:00",
    "modified_at": "2026-10-10T03:43:44.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-09-004-001",
        "text": "reViT matched a full-depth encoder with about 70% fewer stored parameters, while keeping inference computation comparable.",
        "source_ids": [
          "source-2026-10-09-004"
        ],
        "qualification": "reViT restores that variation by blending a small shared bank of feed-forward experts according to a continuous depth coordinate."
      }
    ],
    "source_ids": [
      "source-2026-10-09-004"
    ],
    "tags": [
      "vision transformers",
      "recurrent depth",
      "parameter sharing"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-09-004",
      "title": "arXiv preprint 2610.12448, original v1",
      "publisher": "arXiv / submitting research authors",
      "url": "https://arxiv.org/abs/2610.12448v1",
      "canonical_url": "https://arxiv.org/abs/2610.12448v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-08T13:58:36.000-04:00",
      "accessed_at": "2026-10-10T03:27:42.309-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-09-004-001"
      ]
    }
  ],
  "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 Vision Block Reused Itself at Different Depths",
    "publisher": "The Machine Press",
    "published_at": "2026-10-09T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-09-004/one-vision-block-reused-itself-at-different-depths"
  }
}