Research map / Architectures and efficiency / Quantization
BATQuant: Outlier-Resilient MXFP4 Quantization via Learnable Block-wise Optimization
Jifu Li, Manyi Zhang, Xiaobo Xia, Han Bao, Haoli Bai, Zhenhua Dong, Xianzhi Yu
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 affine transformation instead of global (ablation) | BATQuant · Qwen3-8B W4A4KV16 non-reasoning avg | average accuracy: 68.24 -> 68.70 | Improved |
| use block-wise affine transformation instead of global (ablation) | BATQuant · Qwen3-VL-8B-Instruct W4A4KV16 | recovery rate (%): 95.59 -> 96.43 | Improved |
| add block-wise learnable clipping instead of global clipping (ablation) | BATQuant · Qwen3-8B W4A4KV16 non-reasoning avg | average accuracy: 68.51 -> 68.70 | Improved |
| add block-wise learnable clipping instead of global clipping (ablation) | BATQuant · Qwen3-VL-8B-Instruct W4A4KV16 | recovery rate (%): 96.18 -> 96.43 | Improved |
| replace Naive Kronecker with Global and Private Kronecker (GPK) (ablation) | BATQuant · parameter count (N=4096, g=32) | parameters: 10,240 -> 2,112 | Improved |
| set affine block size equal to MXFP block size g=32 instead of smaller (16) or larger (128) (ablation) | BATQuant · Qwen3-8B and Qwen3-VL-8B | accuracy | Improved |
| vary global shared matrix size g1 (1 to 32) (ablation) | BATQuant GPK · Qwen3-8B and Qwen3-VL-8B | accuracy | Mixed |
| replace FlatQuant with BATQuant | Qwen3-VL-8B-Instruct · W4A4KV16 multimodal | recovery rate (%): 94.79 -> 96.43 | Improved |
| replace GPTQ with BATQuant | Qwen3-8B · W4A8KV16 reasoning | recovery rate (%): 96.54 -> 97.46 | Improved |
| replace GPTQ with BATQuant | Qwen3-8B · W4A4KV16 reasoning | recovery rate (%): 90.10 -> 92.45 | Improved |
| replace GPTQ with BATQuant | Qwen3-8B · W4A8KV8 reasoning | recovery rate (%): 94.54 -> 96.22 | Improved |
| replace rotation-based QuaRot with BATQuant | Qwen3-VL-8B-Instruct · W4A4KV16 multimodal | recovery rate (%): 89.69 -> 96.43 | 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.
- FlatQuant — FlatQuant: Flatness Matters for LLM Quantization (ICML 2025) · better on 6 of 6 shared results
- SmoothQuant — SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models (ICML 2023) · better on 2 of 2 shared results
- SpinQuant — SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025) · better on 1 of 1 shared result
- QuaRot — QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (NeurIPS 2024) · 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
- DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (NeurIPS 2024 · similarity 0.89)
- AffineQuant: Affine Transformation Quantization for Large Language Models (ICLR 2024 · similarity 0.89)
- RaZeR: Pushing the Limits of NVFP4 Quantization with Redundant Zero Remapping (NeurIPS 2026 · similarity 0.88)
- BlockDialect: Block-wise Fine-grained Mixed Format Quantization for Energy-Efficient LLM Inference (ICML 2025 · similarity 0.88)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.88)
- DartQuant: Efficient Rotational Distribution Calibration for LLM Quantization (NeurIPS 2025 · similarity 0.88)
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.
