Compactor: Calibrated Query-Agnostic KV Cache Compression with Approximate Leverage Scores

Abstract: Modern LLMs are increasingly trained to support very large context windows. Unfortunately the ability to use long contexts in generation is complicated by the large memory requirement of the KV cache, which scales linearly with the context length. This memory footprint is often the dominant resource bottleneck in real-world deployments, limiting throughput and increasing serving cost. One way to address this is by compressing the KV cache, which can be done either with knowledge of the question being asked (query-aware) or without knowledge of the query (query-agnostic). We present Compactor, a parameter-free, query-agnostic KV compression strategy that uses approximate leverage scores to determine token importance. We show that Compactor can achieve the same performance as competing methods while retaining 1/2 the tokens in both synthetic and real-world context tasks, with minimal computational overhead. We further introduce a procedure for context-calibrated compression, which allows one to infer the maximum compression ratio a given context can support. Using context-calibrated compression, we show that Compactor achieves full KV performance on Longbench while reducing the KV memory burden by 63%, on average. To demonstrate the efficacy and generalizability of our approach, we apply Compactor to 27 synthetic and real-world tasks from RULER and Longbench, with models from both the Qwen 2.5 and Llama 3.1 families.