IFM Talk: How Discrete Diffusion Speeds Up LLMs w/ Subham Sahoo

​About Event

​🔬 AI4Science on alphaXiv
🗓 Friday November 13th 2026 · 10 AM PT
🎙 Featuring Subham Sahoo's work done w/ IFM
💬 Casual Talk + Open Discussion

​🎥 Zoom: Upon Registration

Description: Most LLMs generate one token at a time. Uno, new work from IFM, uses discrete diffusion to generate multiple tokens in parallel, delivering up to 3× faster generation while preserving the base model’s output distribution.

​Join first author Subham Sahoo for a walkthrough of how Uno combines autoregressive and diffusion modeling in a single model, without needing a separate draft model. He’ll explain how the method can accelerate existing open-weight LLMs, why its speedups persist at large batch sizes, and how it compares with speculative decoding and diffusion language models.

​Bring your questions for a casual talk and open discussion.

Check out the paper here! https://www.alphaxiv.org/abs/2609.04010

Whether you’re working on the frontier of LLMs or just curious about anything AI4Science, we’d love to have you there.

​Hosted by alphaXiv