Saved in:
Bibliographic Details
Main Authors: Ma, Qingsen, Wang, Dianyun, Lyu, Jiaming, Wang, Yaoye, Ning, Lechen, Zhu, Sujie, Xu, Zhenbo, Xiang, Liuyu, Li, Huining, Wu, Huijia, He, Zhaofeng
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2512.10547
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908705293860864
author Ma, Qingsen
Wang, Dianyun
Lyu, Jiaming
Wang, Yaoye
Ning, Lechen
Zhu, Sujie
Xu, Zhenbo
Xiang, Liuyu
Li, Huining
Wu, Huijia
He, Zhaofeng
author_facet Ma, Qingsen
Wang, Dianyun
Lyu, Jiaming
Wang, Yaoye
Ning, Lechen
Zhu, Sujie
Xu, Zhenbo
Xiang, Liuyu
Li, Huining
Wu, Huijia
He, Zhaofeng
contents The Key-Value (KV) cache is the primary memory bottleneck in long-context Large Language Models, yet it is typically treated as an opaque numerical tensor. In this work, we propose \textbf{STA-Attention}, a framework that utilizes Top-K Sparse Autoencoders (SAEs) to decompose the KV cache into interpretable ``semantic atoms.'' Unlike standard $L_1$-regularized SAEs, our Top-K approach eliminates shrinkage bias, preserving the precise dot-product geometry required for attention. Our analysis uncovers a fundamental \textbf{Key-Value Asymmetry}: while Key vectors serve as highly sparse routers dominated by a ``Semantic Elbow,'' deep Value vectors carry dense content payloads requiring a larger budget. Based on this structure, we introduce a Dual-Budget Strategy that selectively preserves the most informative semantic components while filtering representational noise. Experiments on Yi-6B, Mistral-7B, Qwen2.5-32B, and others show that our semantic reconstructions maintain perplexity and zero-shot performance comparable to the original models, effectively bridging the gap between mechanistic interpretability and faithful attention modeling.
format Preprint
id arxiv_https___arxiv_org_abs_2512_10547
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Unlocking the Address Book: Dissecting the Sparse Semantic Structure of LLM Key-Value Caches via Sparse Autoencoders
Ma, Qingsen
Wang, Dianyun
Lyu, Jiaming
Wang, Yaoye
Ning, Lechen
Zhu, Sujie
Xu, Zhenbo
Xiang, Liuyu
Li, Huining
Wu, Huijia
He, Zhaofeng
Machine Learning
The Key-Value (KV) cache is the primary memory bottleneck in long-context Large Language Models, yet it is typically treated as an opaque numerical tensor. In this work, we propose \textbf{STA-Attention}, a framework that utilizes Top-K Sparse Autoencoders (SAEs) to decompose the KV cache into interpretable ``semantic atoms.'' Unlike standard $L_1$-regularized SAEs, our Top-K approach eliminates shrinkage bias, preserving the precise dot-product geometry required for attention. Our analysis uncovers a fundamental \textbf{Key-Value Asymmetry}: while Key vectors serve as highly sparse routers dominated by a ``Semantic Elbow,'' deep Value vectors carry dense content payloads requiring a larger budget. Based on this structure, we introduce a Dual-Budget Strategy that selectively preserves the most informative semantic components while filtering representational noise. Experiments on Yi-6B, Mistral-7B, Qwen2.5-32B, and others show that our semantic reconstructions maintain perplexity and zero-shot performance comparable to the original models, effectively bridging the gap between mechanistic interpretability and faithful attention modeling.
title Unlocking the Address Book: Dissecting the Sparse Semantic Structure of LLM Key-Value Caches via Sparse Autoencoders
topic Machine Learning
url https://arxiv.org/abs/2512.10547