Tsinghua Huang Gao team wins ICML 2026 Outstanding Paper Award, Test of Time Award given to classic algorithm A3C

News on July 6: The International Conference on Machine Learning (ICML) 2026, a top-tier conference in machine learning, was held in Seoul, South Korea, and announced the award-winning papers of the year. The paper co-authored by Tsinghua University's Huang Gao team and Alibaba (The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models) won the Outstanding Paper Award. The research reveals that the flexibility of arbitrary generation order in diffusion language models actually limits model potential in general reasoning tasks such as mathematics and programming, while abandoning arbitrary order and adopting traditional left-to-right generation not only simplifies the method but also significantly improves reasoning accuracy.

Another Outstanding Paper Award was won by MIT and Yale University, whose research proposed a high-accuracy sampling algorithm for diffusion models (High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions), achieving an exponential improvement in the number of steps (or sampling complexity) required to reach a target sampling accuracy.

This year's conference also featured a Position Paper (Position: The Alignment Community is Unintentionally Building a Censor's Toolkit) awarded a major prize, co-authored by researchers from the University of Munich in Germany and independent researchers, pointing out that current AI alignment techniques carry dual-use risks and can easily be maliciously manipulated to become tools of information censorship.

The Test of Time Award at this year's conference was given to Google DeepMind's classic reinforcement learning algorithm published in 2016 (Asynchronous Methods for Deep Reinforcement Learning). The asynchronous advantage actor-critic (A3C) architecture proposed in this study significantly improved the training efficiency of deep reinforcement learning, ushering in an era of efficiently training agents using ordinary multi-core CPUs.