On July 6, 2026, the paper crafted under the guidance of Huang Gao's team from the Automation Department at Tsinghua University, was bestowed with the prestigious Outstanding Paper Award at the 43rd International Conference on Machine Learning (ICML). This paper, a joint endeavor between Tsinghua University and Alibaba, introduces a counterintuitive viewpoint regarding the 'arbitrary order generation' trait of diffusion large language models (dLLMs). Specifically, in general reasoning tasks, such as mathematics and programming, this flexibility can, in fact, diminish the model's reasoning capabilities. The study reveals that the model leverages the freedom of generation order to intentionally sidestep content with high uncertainty, which is crucial for exploration, resulting in a premature collapse of the reasoning space.

To counteract this issue, the team proposed the 'JustGRPO' training method. This approach enforces a left-to-right order generation during the reinforcement learning phase while preserving the advantage of arbitrary order generation during the inference phase. Experimental results demonstrate that after training with JustGRPO, the model attained an accuracy rate of 89.1% on the GSM8K test set, significantly outperforming existing post-training methods for diffusion language models. This paper not only challenges the prevailing assumptions in the field of diffusion language models but also paves a new path for future research endeavors.