Decomposing MXFP4 quantization error for LLM reinforcement learning: reducible bias, recoverable deadzone, and an irreducible floor
The paper discusses the MXFP4 quantization error in reinforcement learning for large language models. It identifies three distinct components of this error and their impact on training. Targeted corrections are proposed to mitigate these issues and recover accuracy.
- ▪MXFP4 arithmetic can significantly speed up reinforcement learning post-training for large language models.
- ▪The quantization error is decomposed into three components: scale bias, deadzone truncation, and grid noise.
- ▪Targeted corrections can recover BF16 accuracy to within 0.7% and 3.0% for specific models.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20402 (cs) [Submitted on 19 May 2026] Title:Decomposing MXFP4 quantization error for LLM reinforcement learning: reducible bias, recoverable deadzone, and an irreducible floor Authors:Xiaocan Li, Shiliang Wu, Zheng Shen View a PDF of the paper titled Decomposing MXFP4 quantization error for LLM reinforcement learning: reducible bias, recoverable deadzone, and an irreducible floor, by Xiaocan Li and 2 other authors View PDF HTML (experimental) Abstract:MXFP4 arithmetic can dramatically accelerate reinforcement learning (RL) post-training of large language models (LLMs), yet the quantization error introduces severe accuracy degradation.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
