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

2026-06-02

AI Summary

This work addresses the challenge of error accumulation in KV-cache quantization during long-sequence autoregressive inference, where outlier token scales severely degrade performance over time steps. The authors propose the first calibration-free KV-cache quantization method tailored for autoregressive decoding, leveraging Hadamard rotation and dual-axis variance normalization of the key and value matrices to effectively suppress quantization error propagation. Supporting 2-bit precision, the method is seamlessly integrated into the vLLM framework and achieves state-of-the-art results on generative benchmarks—including MATH500, AIME24, and HumanEval—at 2-bit accuracy, substantially outperforming existing baselines.

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

Problem

• KV-cache quantization
• error accumulation
• autoregressive decoding
• token-scale errors
• long-horizon reasoning

Innovation

• KV-cache quantization
• variance normalization
• error accumulation
• Hadamard rotation
• calibration-free