{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-31-027",
    "source_story_id": "tmp-story-kv-cache-eviction-front-rail",
    "edition_id": "mp-2026-08-31-morning-0053",
    "edition_url": "https://themachinepress.com/edition/2026-08-31",
    "position": 16,
    "story_type": "dispatch",
    "section": "infrastructure",
    "editorial_classification": "editorial",
    "headline": "KV Cache Eviction Became an Estimation Problem",
    "slug": "kv-cache-eviction-became-an-estimation-problem",
    "dek": "A probabilistic formulation exposes the bias in common eviction heuristics and enables decode-time correction.",
    "summary": "A probabilistic formulation exposes the bias in common eviction heuristics and enables decode-time correction.",
    "body_text": "The paper formalizes optimal KV-cache eviction, proves the general problem computationally hard and recasts practical eviction as expectation estimation. From that view, existing methods become zero-variance biased estimators that can be modified to correct for discarded entries during decoding. The probabilistic variants were more robust across tasks at the same compression budget in the reported experiments, though the formulation does not remove the underlying computational hardness.",
    "why_it_matters": "A probabilistic formulation exposes the bias in common eviction heuristics and enables decode-time correction.",
    "limitations": [
      "The probabilistic variants were more robust across tasks at the same compression budget in the reported experiments, though the formulation does not remove the underlying computational hardness."
    ],
    "importance": 7,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-31-027/kv-cache-eviction-became-an-estimation-problem",
    "json_url": "https://themachinepress.com/story/mp-2026-08-31-027.json",
    "first_published_at": "2026-08-31T09:00:00.000-04:00",
    "modified_at": "2026-08-31T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-08-31-027-001",
        "text": "A probabilistic formulation exposes the bias in common eviction heuristics and enables decode-time correction.",
        "source_ids": [
          "source-2026-08-31-016"
        ],
        "qualification": "The probabilistic variants were more robust across tasks at the same compression budget in the reported experiments, though the formulation does not remove the underlying computational hardness."
      }
    ],
    "source_ids": [
      "source-2026-08-31-016"
    ],
    "tags": [
      "KV cache",
      "inference efficiency",
      "probabilistic reasoning"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-31-016",
      "title": "arXiv preprint 2608.28293",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.28293",
      "canonical_url": "https://arxiv.org/abs/2608.28293",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-08-28T08:57:07.000-04:00",
      "accessed_at": "2026-08-31T08:28:26.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-31-027-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "KV Cache Eviction Became an Estimation Problem",
    "publisher": "The Machine Press",
    "published_at": "2026-08-31T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-31-027/kv-cache-eviction-became-an-estimation-problem"
  }
}