{ "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": "2609.16617", "title": "Divergence Timing and Cumulative Disagreement under KV-Cache Eviction", "summary": "This paper does not propose a replacement technique; it provides a measurement framework (coupling-based decomposition and residual-branch conditional Monte Carlo) that quantifies how KV-cache eviction policies such as SnapKV and recent-token retention diverge from full-cache generation. Adopting the analysis itself costs nothing at inference but requires instrumented paired trajectories to estimate divergence timing and exposure, which adds engineering complexity without changing serving performance. What can fail is transferability: the finding that 85-90% of mismatch mass comes from post-divergence exposure is an exploratory result on 288 documents and two 7-8B models, so eviction-policy rankings and window conclusions may not hold for other workloads, models, or retention budgets.", "rationale": "The paper is a diagnostic study rather than an adoptable method, and its comparative evidence on SnapKV retention settings is exploratory and confined to two open models, so it informs but does not yet justify changing a production eviction policy.", "technique": "Divergence analysis of KV-cache eviction policies", "family": "cache_kv", "recommendation": "observar", "published": "2026-09-15", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.16617", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "This paper does not propose a replacement technique; it provides a measurement framework (coupling-based decomposition and residual-branch conditional Monte Carlo) that quantifies how KV-cache eviction policies such as SnapKV and recent-token retention diverge from full-cache generation. Adopting the analysis itself costs nothing at inference but requires instrumented paired trajectories to estimate divergence timing and exposure, which adds engineering complexity without changing serving performance. What can fail is transferability: the finding that 85-90% of mismatch mass comes from post-divergence exposure is an exploratory result on 288 documents and two 7-8B models, so eviction-policy rankings and window conclusions may not hold for other workloads, models, or retention budgets.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.16617", "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": [ { "anchor": "S1.E1", "label": "(1)", "section": "1 Introduction", "role": "metric", "latex": "R_{\\mathrm{SMC}}(a)=\\mathbb{E}\\!\\left[\\frac{1}{H}\\sum_{t=1}^{H}\\mathbf{1}\\{X_{t}\\neq Y_{t}^{a}\\}\\right].", "mathml": "RSMC(a)=๐”ผ[1Hโˆ‘t=1H๐Ÿ{Xtโ‰ Yta}].", "context": "We make the trajectory comparison precise through recursive overlapโ€“residual maximal coupling [13, 21]. Reference kernels P and intervention kernels Qa generate paired continuations, each with its original marginal law. Their cumulative positional disagreement is", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.E7", "label": "(7)", "section": "3 From Divergence Timing to Cumulative Disagreement", "role": "proposed_method", "latex": "E_{H} =\\frac{1}{H}\\mathbb{E}\\!\\left[\\sum_{t=1}^{H}\\mathbf{1}\\{\\tau0,\\\\\n0,&E_{H}=0,\\end{cases}\\qquad R_{\\mathrm{SMC}}=O_{H}+E_{H}C_{H}.", "mathml": "EH=1H๐”ผ[โˆ‘t=1H๐Ÿ{ฯ„<t}]=1Hโˆ‘s=1Hps(Hโˆ’s),CH={ฮ H/EH,EH>0,0,EH=0,RSMC=OH+EHCH.", "context": "The equality for ฮ H follows from the same measurable partition and Equation (3). To separate the available time from the rate of mismatch, let", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.E10", "label": "(10)", "section": "3 From Divergence Timing to Cumulative Disagreement", "role": "proposed_method", "latex": "\\{R_{\\mathrm{SMC}}(P,Q):\\mathcal{S}_{L}(P,Q)=(p,\\rho)\\}=[R_{L}(p,\\rho),R_{L}(p,\\rho)+B_{L}(p)].", "mathml": "{RSMCโ€‹(P,Q):๐’ฎLโ€‹(P,Q)=(p,ฯ)}=[RLโ€‹(p,ฯ),RLโ€‹(p,ฯ)+BLโ€‹(p)].", "context": "For every admissible profile on an alphabet with at least two symbols, allowing P,Q to vary over all autoregressive kernels subject to that profile, the compatible risks are exactly", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-10-15", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }