Diffusion models are actually worse than traditional autoregressive models in reasoning tasks? This research is quite counterintuitive and worth reading carefully.

The ICML 2026 Outstanding Paper Award was awarded to a paper by Tsinghua University's Huang Gao team in collaboration with Alibaba, pointing out that the flexibility of diffusion language models for arbitrary-order generation can limit their potential in reasoning tasks such as mathematics and programming, while left-to-right generation is more concise and significantly improves reasoning accuracy. The Test of Time Award was awarded to a 2016 paper by Google DeepMind, which proposed the Asynchronous Advantage Actor-Critic (A3C) architecture, greatly improving the training efficiency of deep reinforcement learning.