Masked Diffusion Language Models (MDLMs) have been hindered by the difficulty of estimating log-likelihood in reinforcement learning applications. To address this challenge, researchers have introduced Mask-Aware Policy Gradients, a novel approach that accounts for the sequence of unmasked positions during text generation. By incorporating this positional information, the new method improves the accuracy of log-likelihood estimation, enabling more effective reinforcement learning for MDLMs. This breakthrough has significant implications for the development of large language models, as it enhances their reasoning capabilities. The introduction of Mask-Aware Policy Gradients also raises important security considerations, as the increased power of these models can be exploited for malicious purposes1. As a result, practitioners must carefully evaluate the potential risks and benefits of deploying these advanced language models, particularly in high-stakes applications where security is a top priority.