Probing the Prompt KV Cache: Where It Becomes Dispensable
This research provides insights into the structural redundancy of prompt KV caches for large language models, which could lead to more efficient inference for practitioners.
This paper investigates the redundancy within the prompt KV cache during decoding, specifically identifying when and where it can be replaced without significant accuracy loss. They found that replacing the upper layer prompt span KV cache with a chat template scaffold (using neutral filler content) recovers near clean accuracy, while simply zeroing these slots collapses accuracy.
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.