Research map / Architectures and efficiency / Quantization
BRECQ: Pushing the Limit of Post-Training Quantization by Block Reconstruction
Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, Shi Gu
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 block-wise rather than layer-wise reconstruction (ablation) | BRECQ · ImageNet, ResNet-18, 2-bit weights | top-1 accuracy: 65.19 -> 66.39 | Improved |
| Use block-wise rather than stage-wise reconstruction (ablation) | BRECQ · ImageNet, ResNet-18, 2-bit weights | top-1 accuracy: 66.01 -> 66.39 | Improved |
| Use block-wise rather than network-wise reconstruction (ablation) | BRECQ · ImageNet, ResNet-18, 2-bit weights | top-1 accuracy: 54.15 -> 66.39 | Improved |
| Use block-wise rather than layer-wise reconstruction (ablation) | BRECQ · ImageNet, MobileNetV2, 2-bit weights | top-1 accuracy: 52.13 -> 59.67 | Improved |
| Use block-wise rather than stage-wise reconstruction (ablation) | BRECQ · ImageNet, MobileNetV2, 2-bit weights | top-1 accuracy: 54.23 -> 59.67 | Improved |
| Use block-wise rather than network-wise reconstruction (ablation) | BRECQ · ImageNet, MobileNetV2, 2-bit weights | top-1 accuracy: 40.76 -> 59.67 | Improved |
| Apply 2-bit weight quantization with block reconstruction | BRECQ · ImageNet, ResNet-18 | top-1 accuracy: 55.96 -> 66.30 | Improved |
| Apply 2-bit weight quantization with block reconstruction | BRECQ · ImageNet, MobileNetV2 | top-1 accuracy: 32.54 -> 59.67 | Improved |
| Apply 2-bit weights and 4-bit activations with block reconstruction | BRECQ · ImageNet, ResNet-18 | top-1 accuracy: 0.21 -> 64.80 | Improved |
| Apply 2-bit weights and 4-bit activations with block reconstruction | BRECQ · ImageNet, MobileNetV2 | top-1 accuracy: 0.10 -> 53.34 | Improved |
| Use 4-bit post-training quantization instead of QAT | BRECQ versus PACT · ImageNet, MobileNetV2, 4/4 | top-1 accuracy: 61.40 -> 66.57 | Improved |
| Use 4-bit post-training quantization instead of QAT | BRECQ versus PACT · ImageNet, ResNet-18, 4/4 | GPU hours: 100 -> 0.4 | Improved |
Later papers that compared with this method
16 later accepted papers report BRECQ as a baseline in their results tables; 13 of them report a better number on most of the results they share with it.
- Shift-and-Sum Quantization for Visual Autoregressive Models (ICLR 2026 · better on 7 of 8 shared results)
- LogART: Pushing the Limit of Efficient Logarithmic Post-Training Quantization (ICLR 2026 · better on 3 of 3 shared results)
- Post-Training Quantization for Video Matting (ICLR 2026 · better on 3 of 3 shared results)
- SPA-Q: Structure-Preserving Adaptive Post-Training Quantization for Monocular Depth Estimation (NeurIPS 2026 · better on 2 of 2 shared results)
- QuartDepth: Post-Training Quantization for Real-Time Depth Estimation on the Edge (CVPR 2025 · better on 8 of 8 shared results)
- On Quantizing Neural Representation for Variable-Rate Video Coding (ICLR 2025 · better on 3 of 3 shared results)
- AHCPTQ: Accurate and Hardware-Compatible Post-Training Quantization for Segment Anything Model (ICCV 2025 · better on 3 of 3 shared results)
- Towards Next-Level Post-Training Quantization of Hyper-Scale Transformers (NeurIPS 2024 · better on 0 of 1 shared result)
8 more not listed.
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
- Pushing the Limit of Post-Training Quantization (TPAMI 2025 · similarity 0.93)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · similarity 0.90)
- QDrop: Randomly Dropping Quantization for Extremely Low-bit Post-Training Quantization (ICLR 2022 · similarity 0.89)
- Reg-PTQ: Regression-specialized Post-training Quantization for Fully Quantized Object Detector (CVPR 2024 · similarity 0.88)
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.88)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.87)
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.
