{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-07-11-002",
    "source_story_id": "tmp-feature-million-token-attention",
    "edition_id": "mp-2026-07-11-morning-0002",
    "edition_url": "https://themachinepress.com/edition/2026-07-11",
    "position": 2,
    "story_type": "secondary",
    "section": "research",
    "editorial_classification": "editorial",
    "headline": "Million-Token Memory Without the Global Bottleneck",
    "slug": "million-token-memory-without-the-global-bottleneck",
    "dek": "Microsoft researchers propose interleaving context across GPUs so each device can select locally while the combined result nearly covers the global top tokens.",
    "summary": "Microsoft researchers propose interleaving context across GPUs so each device can select locally while the combined result nearly covers the global top tokens.",
    "body_text": "Exact top-k sparse attention can preserve quality over long contexts, but coordinating a global selection across many GPUs can erase the computational savings. In a July ICML workshop preprint, Microsoft researchers Yifan Guo, Wei Cui, and Peng Cheng propose Interleaved DeepSeek Sparse Attention. Tokens are distributed in an interleaved layout and each device performs a relaxed local selection; the union of those selections is designed to cover nearly all globally relevant tokens with less synchronization. This is a research proposal, not a deployed product claim, and its practical value depends on evaluation across models and hardware. Still, it reframes the long-context problem as a systems question: how to retain useful memory without making every processor wait for a perfect global vote.",
    "why_it_matters": "Microsoft researchers propose interleaving context across GPUs so each device can select locally while the combined result nearly covers the global top tokens.",
    "limitations": [
      "This is a research proposal, not a deployed product claim, and its practical value depends on evaluation across models and hardware."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-07-11-002/million-token-memory-without-the-global-bottleneck",
    "json_url": "https://themachinepress.com/story/mp-2026-07-11-002.json",
    "first_published_at": "2026-07-11T09:00:00.000-04:00",
    "modified_at": "2026-07-11T23:17:20.883-04:00",
    "content_status": "updated",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-07-11-002-001",
        "text": "Microsoft researchers propose interleaving context across GPUs so each device can select locally while the combined result nearly covers the global top tokens.",
        "source_ids": [
          "source-2026-07-11-002"
        ],
        "qualification": "This is a research proposal, not a deployed product claim, and its practical value depends on evaluation across models and hardware."
      }
    ],
    "source_ids": [
      "source-2026-07-11-002"
    ],
    "tags": [
      "long context",
      "sparse attention",
      "distributed inference",
      "ICML 2026",
      "Microsoft Research"
    ],
    "image_url": "https://themachinepress.com/issues/2026-07-11/feature.png",
    "corrections": [
      {
        "correction_id": "correction-2026-07-11-001",
        "revision": 2,
        "summary": "Correction, July 11, 2026: An earlier automated version repeated five ticker briefs and nine dispatches from the July 10 edition. Those items were removed and replaced with previously unpublished editorial stories; the two front-page stories and three genuinely new dispatches were preserved.",
        "published_at": "2026-07-11T23:17:20.883-04:00",
        "affected_story_ids": [],
        "affected_claim_ids": []
      }
    ]
  },
  "sources": [
    {
      "source_id": "source-2026-07-11-002",
      "title": "Microsoft Research: Perfect Recall, Parallel Efficiency",
      "publisher": "Microsoft Research",
      "url": "https://www.microsoft.com/en-us/research/publication/perfect-recall-parallel-efficiency-multi-head-latent-attention-for-million-token-context-decoding/",
      "canonical_url": "https://www.microsoft.com/en-us/research/publication/perfect-recall-parallel-efficiency-multi-head-latent-attention-for-million-token-context-decoding/",
      "source_type": "secondary_reporting",
      "is_primary_source": false,
      "published_at": "2026-07-09T20:00:00.000-04:00",
      "accessed_at": "2026-07-11T08:24:35.519-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-07-11-002-001"
      ]
    }
  ],
  "corrections": [
    {
      "correction_id": "correction-2026-07-11-001",
      "revision": 2,
      "summary": "Correction, July 11, 2026: An earlier automated version repeated five ticker briefs and nine dispatches from the July 10 edition. Those items were removed and replaced with previously unpublished editorial stories; the two front-page stories and three genuinely new dispatches were preserved.",
      "published_at": "2026-07-11T23:17:20.883-04:00",
      "affected_story_ids": [],
      "affected_claim_ids": []
    }
  ],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Million-Token Memory Without the Global Bottleneck",
    "publisher": "The Machine Press",
    "published_at": "2026-07-11T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-07-11-002/million-token-memory-without-the-global-bottleneck"
  }
}