English

Semantic Integrity Matters: Benchmarking and Preserving High-Density Reasoning in KV Cache Compression

Computation and Language 2026-05-13 v4 Artificial Intelligence

Abstract

While Key-Value (KV) cache compression is essential for efficient LLM inference, current evaluations disproportionately focus on sparse retrieval tasks, potentially masking the degradation of High-Density Reasoning where Chain-of-Thought (CoT) coherence is critical. We introduce KVFundaBench to systematically evaluate this gap, revealing a sharp dichotomy: while retrieval tasks remain robust, reasoning tasks exhibit severe Task-Dependent Degradation under aggressive compression due to disrupted CoT links. Extending our analysis to the DeepSeek-R1 model, we uncover that its specialized attention patterns offer unique insights into the fragility of reasoning chains. Guided by these findings -- specifically the necessity of preserving few-shot examples as indivisible Semantic Units -- we propose ShotKV. This approach explicitly separates prefill and decoding phases to prioritize semantic integrity. Empirical results demonstrate that ShotKV achieves 9%-18% accuracy improvements on long-context generation tasks and effectively generalizes to document QA, all while delivering an 11% latency reduction compared to full cache inference.

Keywords

Cite

@article{arxiv.2502.01941,
  title  = {Semantic Integrity Matters: Benchmarking and Preserving High-Density Reasoning in KV Cache Compression},
  author = {Xiang Liu and Zhenheng Tang and Hong Chen and Peijie Dong and Zeyu Li and Xiuze Zhou and Bo Li and Xuming Hu and Xiaowen Chu},
  journal= {arXiv preprint arXiv:2502.01941},
  year   = {2026}
}

Comments

ICML 2026

R2 v1 2026-06-28T21:31:31.496Z