A paper co-authored by Sinho Chewi, “High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions,” written with Fan Chen, Constantinos Daskalakis, and Alexander Rakhlin, received an ICML 2026 Outstanding Paper Award, one of only two papers to earn the top honor at this year’s International Conference on Machine Learning. Chewi is an Assistant Professor of Statistics and Data Science at Yale, who earned his PhD in Mathematics and Statistics from MIT in 2023.
A tough selection
The award was not handed out lightly. Program chairs started with 53 candidates, cut them to a shortlist of 22, and passed those to an 11-member committee that reviewed anonymized papers before choosing two Outstanding Papers and five Honorable Mentions. ICML 2026 accepted more than 6,500 papers.

What the paper does
Diffusion models, the technology behind many modern AI image generators, produce output through a long series of small refinement steps. More steps generally means more accuracy, and more computing cost.
According to the award committee, the paper resolves a long-standing theoretical question: whether a sampler can reach a given accuracy in a number of steps that grows only polylogarithmically as the error target shrinks, rather than polynomially, using only gradient information. In plain terms, the cost of getting much more precise grows very slowly instead of ballooning.
The central tool is a new method the authors call first-order rejection sampling (FORS), which mimics classic rejection sampling without needing the density evaluations earlier high-accuracy methods depended on. The committee described the result as an exponential improvement over prior score-based work.
Why it matters
For diffusion models, the committee noted, the finding suggests that highly accurate sampling could in principle be achieved with far fewer function evaluations. That points toward generative models that are cheaper and faster to run without sacrificing accuracy.
It also fits Chewi’s broader research program. His work has supplied theoretical foundations for Markov chain Monte Carlo methods, variational inference, and score-based diffusion models.
ICML 2026 was held July 6 to 11 in Seoul, South Korea.
