{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-21-002",
    "source_story_id": "tmp-feature-rbs-attention-sparse-prefill",
    "edition_id": "mp-2026-09-21-morning-0074",
    "edition_url": "https://themachinepress.com/edition/2026-09-21",
    "position": 2,
    "story_type": "secondary",
    "section": "chips-infrastructure",
    "editorial_classification": "editorial",
    "headline": "A Sparse Attention Rescue Cut First-Token Time",
    "slug": "a-sparse-attention-rescue-cut-first-token-time",
    "dek": "Radius-bounded selection recovered important tokens that block averages would miss.",
    "summary": "Radius-bounded selection recovered important tokens that block averages would miss.",
    "body_text": "RBS-Attention addresses “mean dilution” in long-context prefill: a block of keys can appear unimportant on average while hiding a strongly relevant token. Its training-free selector combines a centroid branch with a radius-based rescue branch, then runs ordinary block-sparse FlashAttention. On H100 GPUs with a 128,000-token context and Qwen3-30B-A3B, the authors measured a 5.97-fold end-to-end reduction in time to first token against their comparison setup; standalone prefill attention sped up 20.65-fold. A dense Qwen3-32B comparison scored 88.65 on RULER versus 89.52 with dense attention. Those are reported benchmark results for specified models and hardware, not a general speed guarantee.",
    "why_it_matters": "Radius-bounded selection recovered important tokens that block averages would miss.",
    "limitations": [
      "Those are reported benchmark results for specified models and hardware, not a general speed guarantee."
    ],
    "importance": 10,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-21-002/a-sparse-attention-rescue-cut-first-token-time",
    "json_url": "https://themachinepress.com/story/mp-2026-09-21-002.json",
    "first_published_at": "2026-09-21T09:00:00.000-04:00",
    "modified_at": "2026-09-21T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-21-002-001",
        "text": "Radius-bounded selection recovered important tokens that block averages would miss.",
        "source_ids": [
          "source-2026-09-21-002"
        ],
        "qualification": "Those are reported benchmark results for specified models and hardware, not a general speed guarantee."
      }
    ],
    "source_ids": [
      "source-2026-09-21-002"
    ],
    "tags": [
      "long context",
      "sparse attention",
      "inference"
    ],
    "image_url": "https://themachinepress.com/issues/2026-09-21/feature-rbs-attention.png",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-21-002",
      "title": "arXiv preprint 2609.20971",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.20971",
      "canonical_url": "https://arxiv.org/abs/2609.20971",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-17T14:28:29.000-04:00",
      "accessed_at": "2026-09-21T08:26:05.112-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-21-002-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "A Sparse Attention Rescue Cut First-Token Time",
    "publisher": "The Machine Press",
    "published_at": "2026-09-21T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-21-002/a-sparse-attention-rescue-cut-first-token-time"
  }
}