Research map / Architectures and efficiency / Quantization
Bit-Shrinking: Limiting Instantaneous Sharpness for Improving Post-Training Quantization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace direct target-bit quantization with progressive bit-shrinking (ablation) | PTQ models · ImageNet classification | Top-1 accuracy (%): 69.80 -> 69.94; 66.33 -> 67.12; 45.27 -> 57.33 | Helped |
| Use bit-shrinking instead of direct quantization (ablation) | PTQ models · ImageNet classification | Top-1 accuracy (%): 73.81 -> 74.79; 46.77 -> 54.17; 33.31 -> 37.68 | Helped |
| Use bit-shrinking instead of direct quantization (ablation) | PTQ models · ImageNet classification | Top-1 accuracy (%): 68.15 -> 69.02; 32.15 -> 58.66; 2.17 -> 18.23 | Helped |
| Replace integer-bit progressive quantization with adaptive continuous bit shrinking (ablation) | PTQ models · ImageNet classification | Top-1 accuracy (%): 52.41 -> 57.33; 53.15 -> 59.03; 52.58 -> 56.98; 35.49 -> 37.68; 15.32 -> 18.23 | Helped |
| Apply Bit-shrinking rather than QDrop | 4-bit PTQ · ImageNet classification | Top-1 accuracy (%): 69.62 -> 69.94; 75.45 -> 76.04; 68.84 -> 69.02 | Helped |
| Apply Bit-shrinking rather than QDrop | 3-bit PTQ · ImageNet classification | Top-1 accuracy (%): 66.75 -> 67.12; 72.38 -> 72.91; 57.98 -> 58.66 | Helped |
| Apply Bit-shrinking rather than QDrop | 2-bit-weight, 4-bit-activation PTQ · ImageNet classification | Top-1 accuracy (%): 65.25 -> 65.77; 70.65 -> 71.11; 54.22 -> 54.88 | Helped |
| Apply Bit-shrinking rather than QDrop | 2-bit PTQ · ImageNet classification | Top-1 accuracy (%): 54.72 -> 57.33; 58.67 -> 59.03; 13.05 -> 18.23 | Helped |
| Apply Bit-shrinking rather than PTQ4ViT | 8-bit Vision Transformer PTQ · ImageNet-1k validation | Top-1 accuracy (%): 81.00 -> 81.09; 84.25 -> 84.33; 79.47 -> 79.49; 81.48 -> 81.56 | Helped |
| Apply Bit-shrinking rather than PTQ4ViT | 6-bit Vision Transformer PTQ · ImageNet-1k validation | Top-1 accuracy (%): 78.63 -> 80.44; 81.65 -> 83.16; 76.28 -> 78.51; 80.25 -> 80.47; 80.47 -> 80.61; 82.38 -> 82.44 | Mixed |
| Apply Bit-shrinking rather than QDrop | 4-bit object detection PTQ · MS COCO validation | mAP: 33.4 -> 33.8; 37.0 -> 37.4; 32.0 -> 32.5; 35.7 -> 36.0 | Helped |
Closest papers by meaning
- Adaptive Loss-Aware Quantization for Multi-Bit Networks (CVPR 2020 · similarity 0.91)
- Focused Quantization for Sparse CNNs (NeurIPS 2019 · similarity 0.91)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.90)
- And the Bit Goes Down: Revisiting the Quantization of Neural Networks (ICLR 2020 · similarity 0.90)
- Quantization without Tears (CVPR 2025 · similarity 0.89)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (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.