{
  "schema_version": 1,
  "title": "Repro: TileSparse: Arithmetic-Intensity-Aware Sparse Attention for Compute-Bound LLM Decoding",
  "emoji": "🎯",
  "space_id": "dmitry-rov/repro-tilesparse-sparse-attention",
  "paper": {
    "arxiv_id": "",
    "openreview_id": "lK2AzLDJal",
    "icml_id": "9504"
  },
  "tags": [
    "icml2026-repro",
    "paper-lK2AzLDJal"
  ],
  "updated_at": "2026-07-18T08:19:48+00:00",
  "root": {
    "slug": "index",
    "title": "Repro: TileSparse: Arithmetic-Intensity-Aware Sparse Attention for Compute-Bound LLM Decoding",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "claim-1-40-accuracy-improvement-over-dynamic-k-only-sparse-under-tight-compute-budgets",
        "title": "Claim 1: 40% accuracy improvement over dynamic K-only sparse under tight compute budgets",
        "file": "pages/claim-1-40-accuracy-improvement-over-dynamic-k-only-sparse-under-tight-compute-budgets/page.md",
        "children": []
      },
      {
        "slug": "claim-2-99-full-attention-accuracy-at-40-8-attention-compute-reduction",
        "title": "Claim 2: 99% full-attention accuracy at 40.8% attention-compute reduction",
        "file": "pages/claim-2-99-full-attention-accuracy-at-40-8-attention-compute-reduction/page.md",
        "children": []
      },
      {
        "slug": "claim-3-memory-bound-to-compute-bound-shift-with-mla-mtp",
        "title": "Claim 3: memory-bound to compute-bound shift with MLA + MTP",
        "file": "pages/claim-3-memory-bound-to-compute-bound-shift-with-mla-mtp/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      },
      {
        "slug": "tilesparse-repro",
        "title": "tilesparse-repro",
        "file": "pages/tilesparse-repro/page.md",
        "children": []
      }
    ]
  },
  "agent_view_tokens": 7383,
  "revision": "1784362788412981000"
}