Research map / Architectures and efficiency / Quantization
BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Models
Junyu Chen, Jungang Li, Jing Xiong, Wenjie Wang, Qingyao Yang, He Xiao, Zhen Li, Taiqiang Wu, Mengzhao Chen, Zhen Peng, Chaofan Tao, Long Shi and 2 more
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 |
|---|---|---|---|
| use a variable bit-plane grid instead of a fixed uniform grid (BPDQ vs GPTQ) | 2-bit quantization · Qwen2.5-72B GSM8K | accuracy: 63.46% -> 87.72% | Improved |
| use BPDQ instead of GPTQ at 2-bit | W2 quantization · Qwen3-32B GSM8K | accuracy: 44.20% -> 80.89% | Improved |
| use BPDQ instead of GPTQ at 2-bit | W2 quantization · Ministral3-8B GSM8K | accuracy: 12.36% -> 42.46% | Improved |
| use BPDQ instead of AWQ at 2-bit | W2 quantization · Qwen2.5-72B GSM8K | accuracy: 0.00% -> 87.72% | Improved |
| use outlier protection (AWQ) alone at 2-bit | AWQ · 2-bit reasoning | accuracy: 0.00% | Hurt |
| increase group size from 64 to 256 in BPDQ 2-bit (ablation) | BPDQ · Qwen2.5-72B GSM8K | accuracy: 87.72% -> 83.85% | Hurt |
| BPDQ vs VPTQ at 2-bit | BPDQ · Qwen2.5-7B GSM8K | accuracy: 67.63% -> 44.50% | Hurt |
| BPDQ vs GPTQ at 2-bit on small model | BPDQ · Qwen3-0.6B GSM8K | accuracy: 0.00% -> 0.15% | No clear effect |
| BPDQ vs GPTQ at 4-bit | BPDQ · Qwen3-4B GSM8K | accuracy: 85.44% -> 81.50% | Hurt |
| BPDQ vs GPTQ quantization time | BPDQ · Qwen2.5-7B W4 | minutes: 16 -> 47 | Hurt |
| BPDQ vs GPTQ decoding latency 2-bit | BPDQ · Qwen2.5-7B W2 | latency (ms): 33.91 -> 18.09 | Improved |
| fixed-grid GPTQ suppresses outliers at 2-bit | GPTQ-W2-G32 · Qwen2.5-7B | DiagR change: -32.89% | Hurt |
Closest papers by meaning
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.91)
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.90)
- ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals (ICML 2025 · similarity 0.90)
- RSAVQ: Riemannian Sensitivity-Aware Vector Quantization for Large Language Models (NeurIPS 2025 · similarity 0.89)
- Structured Transforms for Low-Overhead Quantization of Language Models (NeurIPS 2026 · similarity 0.89)
- Norm Tweaking: High-Performance Low-Bit Quantization of Large Language Models (AAAI 2024 · 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.
