less than 1 minute read

We have one paper accepted by ACL 2026 Main! This work focuses on efficient long-context reasoning for large language models by reducing the memory and latency overhead of KV cache through query-aware mixed-precision quantization. Detailed information is provided below.

MixKVQ: Query-Aware Mixed-Precision KV Cache Quantization for Long-Context Reasoning

Authors: Tao Zhang, Ziqian Zeng, Hao Peng, Huiping Zhuang, Cen Chen

Abstract:

Long Chain-of-Thought (CoT) reasoning significantly improves the capabilities of large language models, but it also brings substantial memory and latency overhead from the extensive key-value (KV) cache. MixKVQ introduces a lightweight, query-aware mixed-precision KV cache quantization method. It identifies and preserves critical key channels that require higher precision while applying per-token quantization for the value cache, enabling strong long-context reasoning performance with a substantially reduced memory footprint.

Paper: [To be added]

Code: https://github.com/ZeroNLP/MixKVQ

Updated:

You may also enjoy

We have 2 papers accepted by ECCV 2026!

2 minute read

We have two papers accepted by ECCV 2026! The two works advance efficient and trustworthy AI from different perspectives: AnaPFL introduces analytic closed-f...

We have 3 papers accepted by ICLR 2026!

3 minute read

We have three papers accepted by ICLR 2026! One paper advances federated learning with a novel analytic deep framework that tackles data heterogeneity and ef...