{
  "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": "2610.30206",
  "title": "MoEZip: Routing-Aware KV Cache Compression for Sparse Mixture-of-Experts LLMs",
  "summary": "MoEZip replaces routing-agnostic KV eviction methods (e.g., attention-score-only policies) with a scoring system that combines attention weights and a Fisher-information metric of routing sensitivity, so eviction preserves expert selection during generation. The cost is additional per-token scoring computation, implementation complexity beyond standard eviction policies, and applicability restricted to sparse MoE architectures rather than dense models. It can fail if the Fisher-based sensitivity estimate does not transfer to a given MoE model or task, if the serving stack does not expose per-layer routing distributions needed to compute the metric, or if retention ratios below the evaluated regime degrade quality sharply.",
  "rationale": "The method is relevant only if the reader self-hosts MoE models with long-context workloads on the 24 GB GPU; API-based usage cannot apply KV eviction, so adoption requires workload-specific validation against existing eviction baselines.",
  "technique": "Routing-aware KV cache eviction for MoE models",
  "family": "cache_kv",
  "recommendation": "testar",
  "published": "2026-10-21",
  "as_of": "2026-10-24",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2610.30206",
  "independent_implementations": 1,
  "total_implementations": 2,
  "stars": 4,
  "citations": 0,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "MoEZip replaces routing-agnostic KV eviction methods (e.g., attention-score-only policies) with a scoring system that combines attention weights and a Fisher-information metric of routing sensitivity, so eviction preserves expert selection during generation. The cost is additional per-token scoring computation, implementation complexity beyond standard eviction policies, and applicability restricted to sparse MoE architectures rather than dense models. It can fail if the Fisher-based sensitivity estimate does not transfer to a given MoE model or task, if the serving stack does not expose per-layer routing distributions needed to compute the metric, or if retention ratios below the evaluated regime degrade quality sharply.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.30206",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "At 10% KV retention, MoEZip improves over the SoTA baseline by +24 and +20 points on average on LongBench and RULER, corresponding to 2.3x and 4.9x higher scores respectively.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.30206",
      "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": ""
}