🚨 We finally have a better understanding of how the different paradigms for discrete diffusion compare! In new work with Liye Wang, we prove a quadratic separation between the parallel generation capabilities of uniform and Gaussian diffusion relative to unmasking! 1/

Finally here - Submitted PHA-4 to arXiv: drive.google.com/file/d/1m3g… Sampling SK all the way to beta < 1 building on the free-probability and cavity interpolation theory developed in PHA-3, with the amazing @oldheneel and @jtnshi 1/n

There is growing interest in generating discrete data via diffusions. If you have an approach, consider benchmarking it on generating uniform solutions of (random) k-SAT. It is a clean example in which continuous beat discrete diffusions and the diffusion process matters.

If you think no one notices: they do. It's really, really obvious.

📢Aug 31 (Mon): From Interface to Inference: Eliciting Any-Order Inference from Any-Order Models 🤔Many discrete reasoning tasks, such as code generation, are inherently non-causal. Programmers naturally move between high-level structure and local details in a process called any-order inference. While masked diffusion models offer a native any-order prediction interface, maximizing their potential is difficult due to a core limitation: 1️⃣ This any-order interface does not automatically yield actual any-order inference 2️⃣ Fixed-canvas, token-level models suffer from "positional uncertainty"—they may know what semantic component should appear, but not where to place it 🔑To address the interface-inference gap, the authors propose two complementary approaches designed to natively support any-order inference without relying on hand-designed mechanisms. 🔧The framework tackles positional uncertainty through two methods: (1) Insertion-based masked diffusion (building on FlexMDM), which relaxes fixed-position commitments via insertions to enable generation across non-contiguous regions, and (2) Latent-space masked diffusion (LatentMDM), which shifts prediction to coarser semantic segments to enable search over latent generation orders. 📈The resulting models—a 7B FlexMDM trained for Python coding and a 125M LatentMDM for GSM8K—successfully induce distinct any-order inference behaviors. The study demonstrates that overcoming positional uncertainty through these approaches significantly improves downstream performance. This Monday, the joint first authors Seunggeun Kim (@seunggeun_kr), Jaeyeon Kim (@Jaeyeon_Kim_0), Taekyun Lee (taekyunl.github.io), and Yuyuan Chen (yuyuanchen0.github.io) will present their recent paper on the FlexMDM and LatentMDM frameworks.

With the start of the semester just a week away, here’s what I decided to say to my upper division math course about how AI will or won’t change what they’re learning… The Enduring Value of Math, in an Age of AI francissu.substack.com/p/the…

Just because a model can decode tokens in any order doesn’t mean it can truly reason in any order.. Check out Jaeyeon’s awesome thread about our latest work that resolves this conundrum and finally realizes the promise of any-order models!

Rapidly advancing AI models have sparked breakthroughs in long-standing math problems. From solving problems at the International Math Olympiad to disproving decades-old conjectures - some in the math world are now questioning whether they have a future in the field. But not Professor Shayan Oveis Gharan, who won the 2026 Abacus Medal for his work on the theory of algorithms. He believes the power of AI is harnessed not as replacement for human mathematicians, but as a partner. CBS’s Kamal Afzali explains.

Very honestly, the current status of research in this area is more sophisticated than this result. This way of communicating science can make serious and long lasting damage. [This is not to say that the result isn’t new or that AI models aren’t amazing.]

1/ Excited about our new paper w/ @HongYeHu1: Provably Efficient Self-Calibrating Quantum Fault Tolerance. The syndrome bits of QEC code are secretly a calibration signal — and we prove that one can ride them to keep the hardware in tune without ever pausing the computation. ⚛️

Excited to be a part of this amazing community!

A poster session you won't want to miss this afternoon! And stay tuned for some exciting new dLLM results that we have in the works :)

For those lucky enough to be in Seoul right now, make sure to check out Adil's oral presentation in a few hours on our work, joint with Khashayar Gatmiry, on high-accuracy diffusion sampling!