The International Conference on Machine Learning (ICML) 2026, held in Seoul, South Korea, has announced its award-winning papers. A collaboration between Tsinghua University's Huang Gao team and Alibaba took home the Outstanding Paper Award for their work titled The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models. The research reveals that the flexibility of arbitrary generation order in diffusion language models actually limits their potential on reasoning tasks like math and programming. Abandoning arbitrary order in favor of traditional left-to-right generation not only simplifies the method but also significantly improves reasoning accuracy.
Another Outstanding Paper Award went to researchers from MIT and Yale for High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions, which achieves exponential improvements in the number of steps (or sampling complexity) needed to hit a target sampling accuracy.
This year's awards also included a position paper: Position: The Alignment Community is Unintentionally Building a Censor's Toolkit, by researchers from the University of Munich and independent collaborators. It argues that current AI alignment techniques carry dual-use risks and can easily be manipulated into tools for information censorship.
The Test of Time Award was given to Google DeepMind's 2016 classic reinforcement learning paper, Asynchronous Methods for Deep Reinforcement Learning. The work introduced the Asynchronous Advantage Actor-Critic (A3C) architecture, which dramatically improved training efficiency in deep reinforcement learning and enabled efficient training of agents using standard multi-core CPUs.