Saved in:
Bibliographic Details
Main Author: Williams, Jorge L. Ruiz
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.03562
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911698763382784
author Williams, Jorge L. Ruiz
author_facet Williams, Jorge L. Ruiz
contents KV-cache quantizers usually optimize storage-space reconstruction, even though attention reads keys through logits and values through attention-weighted readout. We argue that persistent cache error should be measured in model-visible coordinates. For keys, the visible object is score error modulo constant shifts; this yields HeadQ, a key-side method that stores a low-rank residual side code in a calibration-learned query basis and applies it as an additive logit correction. For values, fixed-attention readout gives an $A^2$-weighted token-distortion surrogate. Across six models, Fisher/score-space error predicts attention KL far better than raw key MSE; same-budget counterexamples, null-space interventions, query-PCA controls, and wrong-sign HeadQ falsify storage-MSE alternatives. Matched Pythia checkpoints localize the main anomaly to a small-model low-entropy route-flip boundary. In K-only WikiText-103 decode experiments with dense values, HeadQ removes roughly $84$--$94\%$ of the excess perplexity on the strongest 2-bit rows; in an auxiliary full-KV 2-bit composition, HeadQ plus an $A^2$ value policy improves all six models.
format Preprint
id arxiv_https___arxiv_org_abs_2605_03562
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle HeadQ: Model-Visible Distortion and Score-Space Correction for KV-Cache Quantization
Williams, Jorge L. Ruiz
Machine Learning
Artificial Intelligence
KV-cache quantizers usually optimize storage-space reconstruction, even though attention reads keys through logits and values through attention-weighted readout. We argue that persistent cache error should be measured in model-visible coordinates. For keys, the visible object is score error modulo constant shifts; this yields HeadQ, a key-side method that stores a low-rank residual side code in a calibration-learned query basis and applies it as an additive logit correction. For values, fixed-attention readout gives an $A^2$-weighted token-distortion surrogate. Across six models, Fisher/score-space error predicts attention KL far better than raw key MSE; same-budget counterexamples, null-space interventions, query-PCA controls, and wrong-sign HeadQ falsify storage-MSE alternatives. Matched Pythia checkpoints localize the main anomaly to a small-model low-entropy route-flip boundary. In K-only WikiText-103 decode experiments with dense values, HeadQ removes roughly $84$--$94\%$ of the excess perplexity on the strongest 2-bit rows; in an auxiliary full-KV 2-bit composition, HeadQ plus an $A^2$ value policy improves all six models.
title HeadQ: Model-Visible Distortion and Score-Space Correction for KV-Cache Quantization
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2605.03562