Trust the Mass: Forced Weights in KV-Cache Eviction

원문 초록

Every deployed KV-cache-eviction rule and every top-$k$ sparse-attention rule keeps a subset of the keys, discards the rest, and renormalizes the attention weights over the kept set. Identifying the exact best subset under that constraint, we find that keeping the largest weights is already near-optimal, since the best subset we find closes only a median $2$ to $5\%$ of the remaining gap to full attention. We evaluate deployed eviction methods in KVPress and find that methods with non-uniform per-head budgets apply attention masks over the full cache because dense cache tensors require equal length across heads. Truncating each head to the target budget lowers benchmark scores by $14$ to $62$ points, and computing token importance with the question in context accounts for $87.6$ points in retrieval accuracy. We propose ContourKV, a training-free rule that applies KVzip's score under a physically enforced budget. It beats KVzip in $93$ of $160$ benchmark runs across RULER and LongBench while using $28$ to $1078$ times less memory, and ties Compactor, the strongest baseline at equal memory.