{
  "schema_version": 1,
  "builder_plan": {
    "scenario": "Long customer conversations exhaust memory or limit concurrent requests.",
    "change": "Change only the attention cache policy, after checking model and runtime support.",
    "baseline": "The same model with its ordinary attention cache.",
    "steps": [
      "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.",
      "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.",
      "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.",
      "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test."
    ],
    "measure": "Peak cache memory, long-context answer quality, first-token latency, and requests completed.",
    "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation."
  },
  "builder_review": null,
  "arxiv_id": "2610.25933",
  "title": "The KV-Surgeon: Compensation-Aware Offline KV-Cache Compaction",
  "summary": "Replaces per-entry independent scoring for KV-cache eviction (e.g., attention matching) with a calibration-data-driven quadratic surrogate that jointly selects removals and compensates remaining entries, accounting for softmax renormalization. Cost is a one-time offline compaction pass with calibration data per fixed context, plus the memory of retaining a small calibration set; the compacted cache then reduces per-query memory and compute on repeated queries. It can fail when the queried context is not fixed and reused (no amortization of the offline cost), when calibration data is unrepresentative of production queries, or when compaction ratios are pushed beyond the validated operating points.",
  "rationale": "The method directly fits constrained single-GPU inference of 4B-30B models and repeated fixed-context workloads, but the gain depends on having stable contexts and suitable calibration data, so workload-specific validation is required before production adoption.",
  "technique": "Compensation-aware offline KV-cache compaction",
  "family": "cache_kv",
  "recommendation": "testar",
  "published": "2026-10-19",
  "as_of": "2026-10-25",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2610.25933",
  "independent_implementations": 2,
  "total_implementations": 2,
  "stars": 0,
  "citations": 0,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "Replaces per-entry independent scoring for KV-cache eviction (e.g., attention matching) with a calibration-data-driven quadratic surrogate that jointly selects removals and compensates remaining entries, accounting for softmax renormalization. Cost is a one-time offline compaction pass with calibration data per fixed context, plus the memory of retaining a small calibration set; the compacted cache then reduces per-query memory and compute on repeated queries. It can fail when the queried context is not fixed and reused (no amortization of the offline cost), when calibration data is unrepresentative of production queries, or when compaction ratios are pushed beyond the validated operating points.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.25933",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "At 50x compaction, improves QA accuracy by up to four points over attention matching (prior SOTA) at comparable cost; a faster variant also exceeds SOTA at two fifths of its cost, evaluated on models from 4B to 30B parameters.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.25933",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "exposure_map": [
    {
      "dimension": "quality",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compute",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "latency",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "operations",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compatibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "security",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "data_and_training",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "reproducibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "risks": [],
  "minimum_test": [],
  "open_questions": [],
  "independent_tests": [],
  "lab_reproductions": [],
  "report_available": false,
  "validation_tier": "unknown",
  "evidence_tier": "unknown",
  "infrastructure_basis": "unknown",
  "equations": [],
  "equations_status": "not_fetched",
  "core_kind": "",
  "equations_fetched_at": "",
  "equations_source_url": "",
  "equations_source_sha256": "",
  "exposures_reviewed_at": "",
  "exposures_source_sha256": ""
}