{ "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": "2610.12133", "title": "Rehearse Everything, Remember Nothing: Attic-KV Rehearses What Will Be Read", "summary": "Replaces full-context rereading as the rehearsal signal for training-free KV cache eviction, instead having the model quiz itself with QA pairs that quote the context plus content-adaptive anchor tokens. It is training-free and compresses faster than full rereading, but adds a rehearsal-generation pass whose cost scales with context and whose self-generated questions must actually quote the relevant spans. It can fail when future queries target content the self-quiz did not rehearse, when the model generates poor or unfaithful quiz questions, or on workloads where answer locations are unpredictable.", "rationale": "It is a training-free, single-GPU-compatible drop-in for KV cache compression with large quality gains at tight budgets, but the rehearsal quality depends on the model's own quiz generation, so the reader should validate it on their specific long-context workload before adopting.", "technique": "Self-quiz rehearsal for KV cache eviction", "family": "cache_kv", "recommendation": "testar", "published": "2026-10-08", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.12133", "independent_implementations": 1, "total_implementations": 2, "stars": 122, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "Replaces full-context rereading as the rehearsal signal for training-free KV cache eviction, instead having the model quiz itself with QA pairs that quote the context plus content-adaptive anchor tokens. It is training-free and compresses faster than full rereading, but adds a rehearsal-generation pass whose cost scales with context and whose self-generated questions must actually quote the relevant spans. It can fail when future queries target content the self-quiz did not rehearse, when the model generates poor or unfaithful quiz questions, or on workloads where answer locations are unpredictable.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.12133", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Best training-free method in all eight tested settings on RULER and LongBench natural-text tasks; up to +41.9 points over full-context rereading at a 3% keep ratio, and +17.1/+28.1 points when plugged into KVgrad and RestoreKV+.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.12133", "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": "" }