Research map / Architectures and efficiency / Attention and transformers
KeepKV: Achieving Periodic Lossless KV Cache Compression for Efficient LLM Inference
Yuxuan Tian, Zihan Wang, Yebo Peng, Aomufei Yuan, Zhiming Wang, Bairen Yi, Xin Liu, Yong Cui, Tong Yang
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 |
|---|---|---|---|
| KeepKV merging versus H2O eviction | Llama-2-7B, 20% KV budget · LongBench NrtvQA | score: 16.54 -> 17.32 | Improved |
| KeepKV versus D2O merging | Llama-2-7B · LongBench TREC | score: 63.5 -> 60.5 | Hurt |
| KeepKV versus D2O merging | Mistral-7B · LongBench HotpotQA | score: 26.41 -> 26.56 | No clear effect |
| KeepKV versus D2O merging | Llama-3-8B · LongBench Qasper | score: 8.86 -> 10.63 | Improved |
| KeepKV versus D2O merging | Llama-3-8B · LongBench RB-P | score: 64.84 -> 65.68 | Improved |
| KeepKV compression versus full cache | Llama-2-7B · throughput at 4k context | tokens/s: 116.54 -> 255.99 | Improved |
| KeepKV versus H2O eviction on throughput | Llama-2-7B · throughput | tokens/s: 317.33 -> 255.99 | Hurt |
| KeepKV versus D2O on throughput | Llama-2-7B · throughput | tokens/s: 214.8 -> 255.99 | Improved |
| combine KeepKV with H2O or PyramidInfer eviction (ablation) | H2O, PyramidInfer · XSUM Llama-7B | ROUGE | Improved |
Closest papers by meaning
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025 · similarity 0.94)
- SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget (ICLR 2025 · similarity 0.94)
- CaM: Cache Merging for Memory-efficient LLMs Inference (ICML 2024 · similarity 0.93)
- EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction (NeurIPS 2026 · similarity 0.93)
- Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference (NeurIPS 2025 · similarity 0.93)
- Get More with LESS: Synthesizing Recurrence with KV Cache Compression for Efficient LLM Inference (ICML 2024 · similarity 0.93)
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.
