MXFP4 quantization error decomposition for LLM reinforcement learning

A recent paper on arXiv (2605.20402) demonstrates that the quantization error of MXFP4 in reinforcement learning for large language models (LLMs) can be broken down into three separate components: scale bias due to power-of-two rounding, deadzone truncation from eliminating small values, and grid noise from rounding to a 4-bit grid. Each of these components leads to specific failure modes: scale bias grows multiplicatively during the backward pass, deadzone truncation reduces the quality of rollouts, and grid noise influences training stability. This analysis indicates that current approaches that view quantization error as a single entity overlook these distinct mechanisms, providing both theoretical and empirical insights into their effects on various RL training pathways.