BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CAIL//Events//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:CAIL Events
BEGIN:VEVENT
UID:2025-10-17-volodymyr-kuleshov@cail.columbia.edu
DTSTAMP:20260823T061016Z
DTSTART:20251017T150000Z
DTEND:20251017T160000Z
SUMMARY:ML Seminar: Volodymyr Kuleshov - Discrete Diffusion Language Model
 s
LOCATION:School of Social Work\, Room C03
DESCRIPTION:While diffusion generative models excel at high-quality image 
 generation\, prior work reports a significant performance gap between diff
 usion and autoregressive (AR) methods on discrete data such as text or bio
 logical sequences. Our work takes steps towards closing this gap via a sim
 ple and effective framework for discrete diffusion. This framework is simp
 le to understand—it optimizes a mixture of denoising (e.g.\, masking) lo
 sses—and can be seen as endowing BERT-like models with principled sample
 rs and variational estimators of log-likelihood. Crucially\, our algorithm
 s are not constrained to generate data sequentially\, and therefore have t
 he potential to improve long-term planning\, controllable generation\, and
  sampling speed.\n\nIn the context of language modeling\, our framework en
 ables deriving masked diffusion language models (MDLMs)\, which achieve a 
 new state-of-the-art among diffusion models\, and approach AR quality. Com
 bined with novel extensions of classifier-free and classifier-based guidan
 ce mechanisms\, these algorithms are also significantly more controllable 
 than AR models. Discrete diffusion extends beyond language to science\, wh
 ere it forms the basis of a new generation of DNA foundation models. Our l
 argest models focus on plants and set a new state of the art in genome ann
 otation\, while also enabling effective generation. Discrete diffusion mod
 els hold the promise to advance progress in generative modeling and its ap
 plications in language understanding and scientific discovery.\n\nhttps://
 cail.columbia.edu/events/2025-10-17-volodymyr-kuleshov
URL:https://cail.columbia.edu/events/2025-10-17-volodymyr-kuleshov
END:VEVENT
END:VCALENDAR
