Value-Aware Stochastic KV Cache Eviction for Reasoning Models

Hugging Face Daily Papers Papers

Summary

VaSE is a training-free method for KV cache eviction that protects large-magnitude value states and introduces stochasticity to improve reasoning model accuracy under compression, outperforming existing methods.

Reasoning models improve accuracy through extended chains of thought, but their long outputs create a memory and compute bottleneck. KV cache eviction methods reduce this cost by evicting unimportant key-value pairs from the cache, yet they often yield worse accuracy than selection-based sparse attention alternatives, which keep the full KV cache. We identify key factors crucial to KV cache eviction accuracy. First, a small fraction of value states have abnormally large magnitudes, and evicting them causes catastrophic failure where models enter repetitive reasoning loops. Second, introducing stochasticity during eviction improves accuracy by increasing cache diversity. Based on these findings, we propose Value-aware Stochastic KV Cache Eviction (VaSE), a training-free recipe that protects large-magnitude value states and promotes diverse eviction decisions. Across six reasoning tasks, Qwen3 models using VaSE with 4x KV cache compression yield higher average accuracies than SOTA selection method at the same sparsity, while outperforming the strongest eviction method by more than 4%. Overall, VaSE bridges the gap between efficiency and accuracy, supporting FlashAttention2 and enabling a static memory footprint for reasoning models.
Original Article
View Cached Full Text

Cached at: 06/03/26, 07:36 AM

Paper page - Value-Aware Stochastic KV Cache Eviction for Reasoning Models

Source: https://huggingface.co/papers/2606.03928

Abstract

Value-aware stochastic KV cache eviction method improves reasoning model accuracy under compression by protecting large-magnitude states and promoting diverse eviction decisions.

Reasoning modelsimprove accuracy through extended chains of thought, but their long outputs create a memory and compute bottleneck.KV cache evictionmethods reduce this cost by evicting unimportant key-value pairs from the cache, yet they often yield worse accuracy than selection-basedsparse attentionalternatives, which keep the full KV cache. We identify key factors crucial toKV cache evictionaccuracy. First, a small fraction ofvalue stateshave abnormally large magnitudes, and evicting them causes catastrophic failure where models enter repetitive reasoning loops. Second, introducingstochasticityduring eviction improves accuracy by increasingcache diversity. Based on these findings, we propose Value-aware StochasticKV Cache Eviction(VaSE), a training-free recipe that protects large-magnitudevalue statesand promotes diverse eviction decisions. Across six reasoning tasks, Qwen3 models using VaSE with 4x KV cache compression yield higher average accuracies than SOTA selection method at the same sparsity, while outperforming the strongest eviction method by more than 4%. Overall, VaSE bridges the gap between efficiency and accuracy, supportingFlashAttention2and enabling a static memory footprint forreasoning models.

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2606\.03928

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.03928 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.03928 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.03928 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Epiphany-Aware KV Cache Eviction Without the Attention Matrix

arXiv cs.LG

This paper introduces EpiKV, a KV cache eviction method that scores token importance via changes in internal representations (epiphany score) instead of attention weights, avoiding the need to materialize the attention matrix. It achieves competitive performance on reasoning benchmarks while enabling up to 16× longer context lengths.