Unlocking the Address Book: Dissecting the Sparse Semantic Structure of LLM Key-Value Caches via Sparse Autoencoders

TL;DR

This paper introduces STA-Attention, a framework using Top-K Sparse Autoencoders to interpret and optimize the Key-Value caches in large language models, revealing semantic structures and improving efficiency.

Contribution

It proposes a novel Top-K SAE method for decomposing KV caches into interpretable semantic components, uncovering key-value asymmetry and enabling selective preservation of important information.

Findings

Semantic reconstructions preserve model perplexity and performance.

Key vectors are sparse routers dominated by a Semantic Elbow.

Value vectors carry dense content requiring larger budgets.

Abstract

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.