Sitan Chen

Excited about this new work where we dig into the role of token order in masked diffusions! MDMs train on some horribly hard tasks, but careful planning at inference can sidestep the hardest ones, dramatically improving over vanilla MDM sampling (e.g. 7%->90% acc on Sudoku) 1/

@SurbhiGoel_ Congrats Surbhi!!

Honored and grateful to be selected as a Sloan fellow this year! A big thanks to my wonderful students, collaborators, and mentors, none of this would be possible without you all.

Excited about this paper where we revisit the core message of our ICML '25 work (diffusion LM training is hard, but enables any-order generation) and develop a new paradigm that achieves 2.5x training speedups by aligning the orders encountered at inference and over training!

📢New paper out! We propose an inference algorithm for diffusion models that does not explicitly depend on the ambient dimension and converges exponentially fast. That’s because, unlike most of the competition, we solve the reverse ODE via Picard and not via Euler discretization

Adam Klivans Wins Test of Time Award at FOCS 2025: cs.utexas.edu/news/2025/adam…

Additionally, please check out the nice concurrent work of Lavenant & Zanella which also proved the connection to Riemann approx of the information curve, plus prior works of Li & Cai and the seminal work of Tim Austin giving operational meaning to dual total correlation. 8/8

Was very fun working with my amazing coauthors Kevin Cong and Jerry Li on this project! Remarkably, Kevin is still an undergrad but could easily pass for a seasoned PhD student given the mathematical level at which he operates.. Paper link: arxiv.org/pdf/2511.04647 7/

Proponents of diffusion language models tout their ability to generate many tokens in parallel. Skeptics argue this is fundamentally broken as it ignores token dependencies. Who's right? 🤔🤔🤔 🚀 In a new work, we rigorously prove that the picture is a lot more nuanced... 1/

Congratulations to the authors for building this awesome resource for the community! Excited to see FlexMDM here 😄

🚨🚨🚨 Now your Masked Diffusion Model can self-correct! We propose PRISM, a plug-and-play approach fine-tuning method that adds self-correction ability to any pretrained MDM! (1/N)

Researchers have demonstrated an algorithm that characterizes quantum systems of any size with optimal efficiency and precision without needing prior information or assumptions about the system’s structure. go.aps.org/4hr1wHO

We found a new way to get language models to reason. 🤯 No RL, no training, no verifiers, no prompting. ❌ With better sampling, base models can achieve single-shot reasoning on par with (or better than!) GRPO while avoiding its characteristic loss in generation diversity.

@RichardKueng @gong_weiyuan Congrats on the beautiful result! This question has been on my mind for a while now, so it’s great to see it finally solved :)

Huge thanks to Viet Tran, Mariami Gachechiladze and MVP Jan Noeller for a productive and fun collaboration! Plus, big shoutout to @sitanch, @gong_weiyuan and Qi Ye for developing elegant new frameworks for proving hardness of learning tasks which we managed to adapt to our needs.

@RobertHuangHY, myself and @preskill identified quantum state learning tasks that look hard, but become easy if you jointly process 2 copies. I have long wondered whether such challenges exist for c>2 copies. Turns out yes, there is an infinite hierarchy: scirate.com/arxiv/2510.080…

Paper link: arxiv.org/abs/2510.08499 6/6

Hard to believe it's been ~5 years since @JordanCotler, @RobertHuangHY, and I started working together on quantum learning under realistic constraints, and while the world looks very different these days, the sheer fun of collaborating w/ them remains a reassuring constant 😀 5/

⚛️⚛️⚛️ Thrilled to share our new paper on quantum probe tomography! In this work we ask: Can one learn about a complex quantum system given only the ability to control and measure a single particle? 1/