The Year in AI Papers
The year/Independent research

Paper 2606.03458

KVarN: Variance-Normalized KV-Cache Quantization Mitigates Error Accumulation in Reasoning Tasks

Lorenz K. Muller, Philippe Bich, Chiara Boretti, Hyun-Min Chang, Jiawei Zhuang, Lukas Cavigelli

Published
Jun 2026
Research lab
Independent
Citations
3
GitHub
452 stars

01 In brief

Summary

The paper introduces KVarN, a calibration-free KV-cache quantization method that mitigates error accumulation during autoregressive decoding in long-horizon reasoning tasks.

It identifies that quantization errors accumulate across timesteps, driven primarily by incorrect token scales, which cause outlier errors that disproportionately degrade end-to-end quality.

KVarN applies a Hadamard rotation followed by dual-scaling variance normalization across both axes of the K and V matrices, fixing token-scale errors and reducing error accumulation.

The method achieves state-of-the-art results on generative benchmarks including MATH500, AIME24, HumanEval, and IFEval at 2-bit precision (2.3 bits per element), with a measured quantization latency overhead of 0.18% over baseline.

A vLLM implementation is available on GitHub.

02 From the paper

Abstract

Test-time scaling is a powerful approach to obtain better reasoning in large language models, but it becomes memory-bottlenecked during long-horizon decoding, as the KV-cache grows. KV-cache quantization can help improve this, but current methods are evaluated under prefill-like settings and errors behave differently under autoregressive decoding. We show that in the latter regime, quantization errors accumulate across timesteps, driven primarily by incorrect token scales. We introduce KVarN, a calibration-free KV-cache quantizer that applies a Hadamard rotation followed by a dual-scaling variance normalization across both axes of the K and V matrices. We find that this combination fixes outlying token-scale errors and substantially reduces error accumulation over existing baselines. KVarN establishes a new state-of-theart for KV-cache quantization on generative benchmarks, including MATH500, AIME24 and HumanEval, at 2-bit precision. A vLLM implementation of the KVarN method is available at https://github.com/huawei-csl/KVarN