Probing the Prompt KV Cache: Where It Becomes Dispensable
Fuente:
arXiv
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866911729600954368 |
|---|---|
| author | Kumar, Vinayshekhar Bannihatti Arivazhagan, Manoj Ghuhan Makhija, Disha Gangadharaiah, Rashmi |
| author_facet | Kumar, Vinayshekhar Bannihatti Arivazhagan, Manoj Ghuhan Makhija, Disha Gangadharaiah, Rashmi |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_30574 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Probing the Prompt KV Cache: Where It Becomes Dispensable Kumar, Vinayshekhar Bannihatti Arivazhagan, Manoj Ghuhan Makhija, Disha Gangadharaiah, Rashmi Computation and Language 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. |
| title | Probing the Prompt KV Cache: Where It Becomes Dispensable |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2605.30574 |