Research map / Architectures and efficiency / Quantization
Unveiling the Potential of Quantization with MXFP4: Strategies for Quantization Error Reduction
Jatin Chhugani, Geonhwa Jeong, Bor-Yiing Su, Yunjie Pan, Hanmei Yang, Aayush Ankit, Jiecao Yu, Summer Deng, Yunqing Chen, Nadathur Satish, Changkyu Kim
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 Overflow-Aware Scaling to block-16 MXFP4 (ablation) | MXFP4-16 on Llama-3.1-8B-Instruct · Downstream average | accuracy: 60.55 -> 63.39 | Improved |
| add Overflow-Aware Scaling to block-16 MXFP4 (ablation) | MXFP4-16 on Qwen3-8B · Downstream average | accuracy: 66.82 -> 69.71 | Improved |
| add static Macro Block Scaling on top of OAS (ablation) | MXFP4-16-OAS on Llama-3.1-8B-Instruct · Downstream average | accuracy: 63.39 -> 65.77 | Improved |
| add static Macro Block Scaling on top of OAS (ablation) | MXFP4-16-OAS on Qwen3-8B · Downstream average | accuracy: 69.71 -> 70.50 | Improved |
| use dynamic MBS for weights (hybrid) instead of static MBS everywhere (ablation) | MXFP4-MBS-S on Llama-3.1-8B-Instruct · Downstream average | accuracy: 65.77 -> 66.50 | Improved |
| use dynamic MBS for weights (hybrid) instead of static MBS everywhere (ablation) | MXFP4-MBS-S on Qwen3-8B · Downstream average | accuracy: 70.50 -> 70.84 | Improved |
| reduce block size from 32 to 16 alone (ablation) | MXFP4-OCP on Llama-3.1-8B-Instruct · Downstream average | accuracy: 61.25 -> 60.55 | Hurt |
| reduce block size from 32 to 16 alone (ablation) | MXFP4-OCP on Qwen3-8B · Downstream average | accuracy: 65.50 -> 66.82 | Improved |
| add OAS to block-16 MXFP4 (ablation) | MXFP4-16 on DeepSeek-R1 · MMLU-Pro | accuracy: 76.29 -> 75.36 | Hurt |
| add OAS to block-16 MXFP4 (ablation) | MXFP4-16 on Llama-3.1-8B-Instruct · Wikitext perplexity | perplexity: 11.52 -> 10.57 | Improved |
| replace OCP MXFP4 with MBS-H | MXFP4-OCP on Qwen3-8B · MMLU-Pro | accuracy: 43.78 -> 59.30 | Improved |
| increase MBS block size from 32 to 512 (ablation) | MBS on Llama-3.1-8B-Instruct · Output QSNR | QSNR (dB): approximately 1.1 dB degradation | Hurt |
Closest papers by meaning
- RaZeR: Pushing the Limits of NVFP4 Quantization with Redundant Zero Remapping (NeurIPS 2026 · similarity 0.91)
- BlockDialect: Block-wise Fine-grained Mixed Format Quantization for Energy-Efficient LLM Inference (ICML 2025 · similarity 0.88)
- Normalized Architectures are Natively 4-Bit (NeurIPS 2026 · similarity 0.88)
- ScaleBITS: Scalable Bitwidth Search for Hardware-Aligned Mixed-Precision LLMs (NeurIPS 2026 · similarity 0.87)
- Optimizing Large Language Model Training Using FP4 Quantization (ICML 2025 · similarity 0.87)
- GANQ: GPU-Adaptive Non-Uniform Quantization for Large Language Models (ICML 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.
