Research map / Architectures and efficiency / Attention and transformers
RaBitQCache: Rotated Binary Quantization for KVCache in Long Context LLM Inference
Wenhao Li, Jinhao Dong, Hailin Zhang, Wenhang Shi, WEI LU, Xiaoyong Du
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 centroid re-centering before quantization (ablation) | RaBitQCache · LongBench LLaMA-3.1-8B | generation score: 50.25 -> 50.63 | Improved |
| use adaptive top-p (p=0.95) vs full attention | RaBitQCache · LongBench LongChat | generation score: 35.78 -> 36.21 | No clear effect |
| use adaptive top-p (p=0.95) vs full attention | RaBitQCache · LongBench LLaMA-3.1-8B | generation score: 50.58 -> 50.6 | No clear effect |
| use RaBitQCache instead of Quest-1024 | LongBench LLaMA-3.1-8B · LongBench LLaMA-3.1-8B | generation score: 46.52 -> 50.6 | Improved |
| use RaBitQCache instead of Double Sparse-256 | LongChat · LongBench LongChat | generation score: 3.75 -> 36.21 | Improved |
| use RaBitQCache vs Quest-4096 | RaBitQCache · LongBench LongChat | generation score: 36.21 -> 36.28 | Hurt |
| use RaBitQCache instead of Quest on 70B | RaBitQCache · LongBench LLaMA-3.1-70B | generation score: 50.69 -> 54.58 | Improved |
| use RaBitQCache instead of Quest on GSM8K | RaBitQCache · GSM8K | accuracy: 0.70 -> 0.77 | Improved |
| use RaBitQCache instead of PyramidKV/SnapKV | RaBitQCache · LongBench LLaMA-3.1-8B | generation score: 45.09 -> 50.63 | Improved |
| use optimized warp kernel vs naive Int4DotBinary (ablation) | Int4DotBinary kernel · operator microbenchmark | speedup: up to 3.43x | Improved |
| use RaBitQCache decoding vs full attention | RaBitQCache · decoding latency 10k-32k | speedup: up to 3.88x | Improved |
| use RaBitQCache end to end vs full precision | RaBitQCache · end-to-end latency | speedup: 2.16x | Improved |
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.
- Quest 256 — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · better on 9 of 9 shared results
- MagicPIG — MagicPIG: LSH Sampling for Efficient LLM Generation (ICLR 2025) · better on 1 of 1 shared result
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
- Efficient Low Rank Attention for Long-Context Inference in Large Language Models (NeurIPS 2025 · similarity 0.91)
- KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled Quantization (NeurIPS 2024 · similarity 0.90)
- RocketKV: Accelerating Long-Context LLM Inference via Two-Stage KV Cache Compression (ICML 2025 · similarity 0.90)
- ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification (NeurIPS 2024 · similarity 0.89)
- Cache Me If You Must: Adaptive Key-Value Quantization for Large Language Models (ICML 2025 · similarity 0.89)
- KVTuner: Sensitivity-Aware Layer-Wise Mixed-Precision KV Cache Quantization for Efficient and Nearly Lossless LLM Inference (ICML 2025 · similarity 0.89)
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.
