{ "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": "2610.03135", "title": "Page-EntroKV: Hardware-Aligned, Entropy-Weighted KV-Cache Eviction under Grouped-Query Attention", "summary": "Page-EntroKV replaces per-query-head KV eviction scoring and mean-pooled head aggregation with entropy-weighted pooling of heads within each GQA group, evicting at the hardware granularity (layer, group, page) that PagedAttention actually allocates. It costs one inner product per head at prefill with no calibration, but requires implementing eviction logic tied to the serving engine's page layout rather than dropping into an off-the-shelf cache policy. It can fail because all results come from head-independent replay on a single 1.5B pilot model, so end-to-end serving behavior, quality on real long-context workloads at larger scale, and integration overhead in engines like vLLM remain unvalidated.", "rationale": "The method directly addresses a real GQA serving inefficiency and is cheap to compute, but the evidence is a single small-model simulation, so the reader should validate it on their own workload before production adoption.", "technique": "GQA-aligned entropy-weighted KV-cache eviction", "family": "cache_kv", "recommendation": "testar", "published": "2026-10-02", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.03135", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "Page-EntroKV replaces per-query-head KV eviction scoring and mean-pooled head aggregation with entropy-weighted pooling of heads within each GQA group, evicting at the hardware granularity (layer, group, page) that PagedAttention actually allocates. It costs one inner product per head at prefill with no calibration, but requires implementing eviction logic tied to the serving engine's page layout rather than dropping into an off-the-shelf cache policy. It can fail because all results come from head-independent replay on a single 1.5B pilot model, so end-to-end serving behavior, quality on real long-context workloads at larger scale, and integration overhead in engines like vLLM remain unvalidated.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.03135", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Holds union overhead ratio at exactly 1.000 versus up to 4.75x cache inflation for head-independent eviction at a 2% budget, with 100% needle recall at 20% retention versus 0% for mean pooling.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.03135", "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": "" }