Probing the Prompt KV Cache: Where It Becomes Dispensable

2026-05-28

AI Summary

This study investigates the redundancy of prompt key-value (KV) cache in large language model inference and the conditions under which it can be substituted. Through controlled KV cache replacement experiments across different network layers and decoding steps on mainstream models—including Qwen3, Gemma 3, and Llama 3—the work reveals for the first time that the redundancy primarily stems from structural scaffolding such as chat templates rather than the semantic content of the prompt itself. The authors propose replacing the original prompt KV cache with a neutrally filled template-based cache, which substantially recovers model performance—achieving accuracy close to the original—whereas zeroing out the cache leads to severe degradation. These findings underscore the critical role of structural information preserved in the KV cache for maintaining generation quality.

Abstract

Prior KV cache compression schemes empirically demonstrate that the prompt cache is partially redundant during decoding, dropping or summarising entries with little accuracy loss. We ask when and what kind of redundancy: at which layers, after how many decoding steps, and in what form can the prompt span KV cache be replaced without breaking the task. A controlled splice intervention swept over layer cutoff and decoding steps shows this redundancy is about form (chat template scaffolding) rather than content. Replacing the upper layer prompt span KV cache with KV cache from a chat template scaffold whose user content is a neutral filler recovers near clean accuracy, while zeroing the same slots collapses accuracy. The dissociation replicates across the Qwen3, Gemma 3, and Llama 3 families on multiple datasets.

Problem

• KV cache
• prompt redundancy
• decoding
• language models
• cache compression

Innovation

• KV cache compression
• prompt redundancy
• chat template scaffolding
• splice intervention
• decoding efficiency