Research map / Architectures and efficiency / Quantization
QuEPT: Quantized Elastic Precision Transformers with One-Shot Calibration for Multi-Bit Switching
Ke Xu, Yixin Wang, Zhongcheng Li, Hao Cui, Jinshui Hu, Xingyi Zhang
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 weight clipping parameters (ablation) | QuEPT · ImageNet ViT-S W4A4 | top-1 accuracy: 69.2 -> 70.1 | Improved |
| add MB-CLoRA to clipping (ablation) | QuEPT · ImageNet ViT-S | top-1 accuracy: W4A4 70.1 -> 72.2; W6A6 80.2 -> 79.2; W8A8 81.1 -> 80.7 | Mixed |
| add MB-ToMe to clipping (ablation) | QuEPT · ImageNet ViT-S W4A4 | top-1 accuracy: 70.1 -> 73.0 | Improved |
| add MB-ToMe on top of MB-CLoRA (ablation) | QuEPT · ImageNet ViT-S W4A4 | top-1 accuracy: 72.2 -> 73.6 | Improved |
| use MAE loss instead of MSE loss (ablation) | QuEPT · ImageNet ViT-S W4A4 | top-1 accuracy: 73.6 -> 74.7 | Improved |
| use cascaded MB-CLoRA instead of independent LoRA (ablation) | QuEPT · LLaMA2-7B W4A4 0-shot avg | accuracy: 59.2 -> 61.6 | Improved |
| use cascaded MB-CLoRA instead of fully shared LoRA (ablation) | QuEPT · LLaMA2-7B W4A4 0-shot avg | accuracy: 60.9 -> 61.6 | Improved |
| use cascaded MB-CLoRA instead of independent LoRA at high bits (ablation) | QuEPT · LLaMA2-7B W7A7/W8A8 | accuracy: 65.7 vs 65.6 | No clear effect |
| use selective merge MB-ToMe instead of random selection (ablation) | QuEPT · LLaMA2-7B W4A4 | 0-shot avg accuracy: 55.7 -> 61.6 | Improved |
| use selective merge MB-ToMe instead of uniform fusion (ablation) | QuEPT · LLaMA2-7B W4A4 | 0-shot avg accuracy: 60.2 -> 61.6 | Improved |
| use selective merge MB-ToMe instead of random selection (ablation) | QuEPT · ImageNet ViT-S W4A4 | top-1 accuracy: 74.5 -> 75.1 | Improved |
| use QuEPT multi-bit instead of PTMQ | QuEPT vs PTMQ · ImageNet ViT-B W6A6/W7A7/W8A8 | top-1 accuracy: +6.1%, +5.6%, +5.2% | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- DuQuant — DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (NeurIPS 2024) · better on 4 of 5 shared results
- PTMQ — PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024) · better on 2 of 2 shared results
- QuaRot — QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (NeurIPS 2024) · better on 1 of 1 shared result
- RepQ-ViT — RepQ-ViT: Scale Reparameterization for Post-Training Quantization of Vision Transformers (ICCV 2023) · 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
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.87)
- The Quantization Benefits of Residual-Free Transformers (NeurIPS 2026 · similarity 0.86)
- QJL: 1-Bit Quantized JL Transform for KV Cache Quantization with Zero Overhead (AAAI 2025 · similarity 0.86)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.86)
- FrameQuant: Flexible Low-Bit Quantization for Transformers (ICML 2024 · similarity 0.86)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.85)
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.
