AvoKV-E: Payload-Aware KV Cache Eviction for Long Reasoning

Long-output reasoning shifts the KV-cache bottleneck from prompt storage to generated reasoning traces. Existing reasoning-cache eviction methods mainly estimate routing utility: whether an old key will still be read, will recur, or can be replaced. We show that this routing-only view misses two effects in long reasoning. First, some low-attention entries carry large value payloads whose removal causes high future KL divergence. Second, newly generated reasoning states can appear stale before later tokens have had time to read them. We propose AvoKV-E, a training-free eviction policy that delays eligibility for newly generated states and ranks eligible entries by read pressure, key redundancy, and value-payload potential after normalizing each signal within the candidate set. In empirical evaluations, AvoKV-E matches or exceeds redundancy-aware, recurrence-based, and thought-adaptive eviction baselines under matched active-KV budgets, with the largest gains in the tightest-cache regime. The results suggest that long-reasoning KV eviction should preserve not only keys that are read, but also value payloads that sustain reasoning.