{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-26-012",
    "source_story_id": "tmp-story-puzzlekv-page-compression",
    "edition_id": "mp-2026-08-26-morning-0048",
    "edition_url": "https://themachinepress.com/edition/2026-08-26",
    "position": 12,
    "story_type": "dispatch",
    "section": "research",
    "editorial_classification": "editorial",
    "headline": "Each Cache Page Found Its Own Low-Rank Basis",
    "slug": "each-cache-page-found-its-own-low-rank-basis",
    "dek": "PuzzleKV compresses completed per-head pages independently instead of sharing one projection across a broad cache region.",
    "summary": "PuzzleKV compresses completed per-head pages independently instead of sharing one projection across a broad cache region.",
    "body_text": "At roughly 60 percent of original KV storage, PuzzleKV retained more than 96 percent of Full KV performance across both evaluated models and all reported settings. Combined with quantization, it retained more than 93 percent using 18.7 percent of storage, with attention computed directly over dense and factorized pages.",
    "why_it_matters": "PuzzleKV compresses completed per-head pages independently instead of sharing one projection across a broad cache region.",
    "limitations": [],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-26-012/each-cache-page-found-its-own-low-rank-basis",
    "json_url": "https://themachinepress.com/story/mp-2026-08-26-012.json",
    "first_published_at": "2026-08-26T09:00:00.000-04:00",
    "modified_at": "2026-08-26T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-08-26-012-001",
        "text": "PuzzleKV compresses completed per-head pages independently instead of sharing one projection across a broad cache region.",
        "source_ids": [
          "source-2026-08-26-012"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-08-26-012"
    ],
    "tags": [
      "KV cache",
      "long context",
      "compression"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-26-012",
      "title": "arXiv preprint 2608.23843",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.23843",
      "canonical_url": "https://arxiv.org/abs/2608.23843",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-08-24T17:30:11.000-04:00",
      "accessed_at": "2026-08-26T08:17:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-26-012-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Each Cache Page Found Its Own Low-Rank Basis",
    "publisher": "The Machine Press",
    "published_at": "2026-08-26T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-26-012/each-cache-page-found-its-own-low-rank-basis"
  }
}