Research map / Architectures and efficiency / Attention and transformers
FreeKV: Boosting KV Cache Retrieval for Efficient LLM Inference
Guangda Liu, Chengwei Li, Zhenyu Ning, Jing Lin, Yiwu Yao, Danning Ke, Minyi Guo, Jieru Zhao
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 |
|---|---|---|---|
| add speculative retrieval using the last step query instead of the last layer query (ablation) | KV retrieval (Qwen DeepSeek-R1-Qwen-7B) · AIME24 | avg@k accuracy: 33.00 -> 50.00 | Improved |
| use last step query instead of last layer query (ablation) | DeepSeek-R1-Qwen-7B · GPQA | avg@k accuracy: 30.50 -> 36.00 | Improved |
| use last step query instead of last layer query (ablation) | Qwen-2.5-7B-Instruct · LongBench v2 | accuracy: 26.44 -> 26.63 | No clear effect |
| add fine-grained correction with tau=0.9 to speculative retrieval (ablation) | DeepSeek-R1-Qwen-7B · AIME24 | avg@k accuracy: 50.00 -> 52.92 | Improved |
| add fine-grained correction with tau=0.9 (ablation) | DeepSeek-R1-Qwen-7B · GPQA | avg@k accuracy: 36.00 -> 39.50 | Improved |
| raise correction threshold from 0.8 to 0.9 (ablation) | DeepSeek-R1-Qwen-7B · GPQA | avg@k accuracy: 35.50 -> 39.50 | Improved |
| use mean pooling over softmax scores (MeanS) instead of MaxS (ablation) | group-consistent selection · AIME24 | avg@k accuracy: 47.50 -> 52.92 | Improved |
| use mean pooling over softmax scores instead of mean query pooling (MeanQ) (ablation) | group-consistent selection · AIME24 | avg@k accuracy: 46.67 -> 52.92 | Improved |
| max instead of mean pooling for group correction (ablation) | group-consistent correction · AIME24 | avg@k accuracy: 52.92 -> 53.33 | No clear effect |
| use FreeKV instead of RazorAttention-style KV dropping | FreeKV · LongGenBench | CR: 35.90 -> 78.03 | Improved |
| use FreeKV instead of ShadowKV on Qwen models | FreeKV · LongGenBench | CR x Acc: 11.43 -> 32.81 | Improved |
| use FreeKV instead of ShadowKV on Llama-3.1-8B | FreeKV · LongGenBench | CR x Acc: 30.66 -> 27.62 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ShadowKV — ShadowKV: KV Cache in Shadows for High-Throughput Long-Context LLM Inference (ICML 2025) · better on 5 of 6 shared results
- Quest — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · better on 6 of 6 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
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.93)
- RocketKV: Accelerating Long-Context LLM Inference via Two-Stage KV Cache Compression (ICML 2025 · similarity 0.92)
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025 · similarity 0.92)
- SpeCache: Speculative Key-Value Caching for Efficient Generation of LLMs (ICML 2025 · similarity 0.92)
- ShadowKV: KV Cache in Shadows for High-Throughput Long-Context LLM Inference (ICML 2025 · similarity 0.92)
- EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction (NeurIPS 2026 · similarity 0.92)
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.
