{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-09-017",
    "source_story_id": "tmp-brief-sparsedecoding-generation-calibration",
    "edition_id": "mp-2026-10-09-morning-0091",
    "edition_url": "https://themachinepress.com/edition/2026-10-09",
    "position": 19,
    "story_type": "ticker",
    "section": "chips-infrastructure",
    "editorial_classification": "editorial",
    "headline": "Pruning Followed the Tokens the Model Actually Generated",
    "slug": "pruning-followed-the-tokens-the-model-actually-generated",
    "dek": "SparseDecoding calibrates on autoregressive activations and adds a sparse matrix-vector kernel.",
    "summary": "SparseDecoding calibrates on autoregressive activations and adds a sparse matrix-vector kernel.",
    "body_text": "Natural-text prefill and self-generated decoding can have different activation distributions. Calibrating on generation, excluding prefill, improved the tested pruned models’ long-form results. An optimized N:M kernel delivered up to 1.48 times end-to-end decoding speedup on A100 GPUs across the evaluated model families.",
    "why_it_matters": "SparseDecoding calibrates on autoregressive activations and adds a sparse matrix-vector kernel.",
    "limitations": [],
    "importance": 7,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-09-017/pruning-followed-the-tokens-the-model-actually-generated",
    "json_url": "https://themachinepress.com/story/mp-2026-10-09-017.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-2026-10-09-017-001",
        "text": "SparseDecoding calibrates on autoregressive activations and adds a sparse matrix-vector kernel.",
        "source_ids": [
          "source-2026-10-09-019"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-10-09-019"
    ],
    "tags": [
      "sparse decoding",
      "inference efficiency"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-09-019",
      "title": "arXiv preprint 2610.12327, original v1",
      "publisher": "arXiv / submitting research authors",
      "url": "https://arxiv.org/abs/2610.12327v1",
      "canonical_url": "https://arxiv.org/abs/2610.12327v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-08T13:06:32.000-04:00",
      "accessed_at": "2026-10-10T03:27:42.309-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-09-017-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Pruning Followed the Tokens the Model Actually Generated",
    "publisher": "The Machine Press",
    "published_at": "2026-10-09T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-09-017/pruning-followed-the-tokens-the-model-actually-generated"
  }
}