Risk-Controlled KV-Cache Eviction: From Memory Budgets to Risk Targets

Long customer conversations exhaust memory or limit concurrent requests.

What you would change

Change only the attention cache policy, after checking model and runtime support.

Change only the attention cache policy, after checking model and runtime support.

Compare against

The same model with its ordinary attention cache.

1. 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.
2. 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.
3. Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.
4. Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test.

What to measure

Peak cache memory, long-context answer quality, first-token latency, and requests completed.

When to keep it

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.

recommendation

test

research area

KV cache

technical change

Risk-controlled KV-cache eviction certification

published

2026-09-23

The certification layer is cheap to implement on a single GPU and workload-specific, but its value depends on whether the reader's deployment actually supports any certified eviction level, which must be validated on their own calibration data.

1. claim 01Paperraft inference
   Read the brief interpretation

   The method replaces ad hoc eviction-budget tuning (choosing a retention level from average quality-memory curves) with a post-hoc, compressor-agnostic certification procedure that picks a retention policy from calibration data under a finite-sample risk guarantee, falling back to full KV when nothing certifies. It costs calibration data drawn from the target workload, an extra utility evaluation against full-KV inference, and potentially large memory headroom: on some workloads no eviction at all is certified, so the safety guarantee can erase the memory savings. It can fail when calibration data is unrepresentative of production traffic, when the utility metric mismeasures task quality, or when the finite-sample correction rejects policies that are actually acceptable, leaving the system at costly full KV.

   This is a Paperraft inference. It is not pinned to a verified PDF passage.

   Inspect the source
2. claim 02Paperraft inference

   On Llama/LongBench, certification selects policies retaining 5-10 percentage points less cache than uncertified empirical choices at a 5% degradation-risk target; on Llama/RULER-32K no compressed policy is certified and full KV is required.

   This is a Paperraft inference. It is not pinned to a verified PDF passage.

   Inspect the source