Research map / Architectures and efficiency / Attention and transformers
ProphetKV: User-Query-Driven Selective Recomputation for Efficient KV Cache Reuse in Retrieval-Augmented Generation
Shihao Wang, Jiahao Chen, Yanqi Pan, Hao Huang, Yichen Hao, Xiangyu Zou, Wen Xia, Chongyang Qiu, Wentao Zhang, PengFei Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select recompute tokens by user-query attention (ProphetKV) vs CacheBlend | KV cache reuse · RULER 8K, Llama3.1-8B-Instruct, 20% recompute | average accuracy: 77.27 -> 84.71 | Improved |
| select recompute tokens by user-query attention vs CacheBlend | KV cache reuse · LongBench, Llama3.1-8B-Instruct, 20% recompute | average accuracy: 39.82 -> 50.80 | Improved |
| select recompute tokens by user-query attention vs CacheBlend | KV cache reuse · LongBench, Qwen2.5-14B-Instruct, 20% recompute | average accuracy: 35.40 -> 53.43 | Improved |
| select recompute tokens by user-query attention vs CacheBlend | KV cache reuse · RULER 8K, Qwen3-14B-Thinking, 20% recompute | average accuracy: 71.99 -> 89.89 | Improved |
| add 20% partial recomputation vs naive reuse | ProphetKV · LongBench, Qwen2.5-14B-Instruct | average accuracy: 15.86 -> 53.43 | Improved |
| fuse query attention across all layers (uniform average) instead of single layer (ablation) | ProphetKV · layer-wise optimal token overlap | overlap ratio | Improved |
| use average layer fusion vs fixed-depth selection (ablation) | ProphetKV · fixed-depth ablation | accuracy | No clear effect |
| use query attention proxy vs CacheBlend under oracle selection (ablation) | ProphetKV · RULER-MV | recompute ratio to converge: 0.06 -> 0.04 | Improved |
| increase recompute ratio (ablation) | ProphetKV · RULER-MV, LongBench Musique | accuracy | Improved |
| decrease chunk size (ablation) | all KV reuse methods · RULER-MV | accuracy | Hurt |
| ProphetKV vs full recomputation latency | ProphetKV · TTFT 16K, Llama3.1-8B-Instruct | TTFT (s): 5.23 -> 1.13 | Improved |
| ProphetKV vs EPIC latency | ProphetKV · TTFT 16K, Llama3.1-8B-Instruct | TTFT (s): 1.08 vs 1.13 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports ProphetKV as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- AgentKVShift: Efficient KV Cache Reuse for Agentic Memory Systems (NeurIPS 2026 · better on 9 of 9 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- KV Packet: Recomputation-Free Context-Independent KV Caching for LLMs (NeurIPS 2026 · similarity 0.90)
- RazorAttention: Efficient KV Cache Compression Through Retrieval Heads (ICLR 2025 · similarity 0.90)
- KVLink: Accelerating Large Language Models via Efficient KV Cache Reuse (NeurIPS 2025 · similarity 0.89)
- KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction (NeurIPS 2025 · similarity 0.88)
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.88)
- ArkVale: Efficient Generative LLM Inference with Recallable Key-Value Eviction (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
