{"693751":{"#nid":"693751","#data":{"type":"event","title":"CSIP Seminar | William Yuan","body":[{"value":"\u003Cp\u003E\u003Cstrong\u003ETalk Title:\u003C\/strong\u003E Understanding and Improving Low-Step Discrete Diffusion Sampling\u003Cbr\u003E\u003Cstrong\u003ESpeaker: \u003C\/strong\u003EWilliam Yuan, Georgia Tech Ph.D. Student\u003Cbr\u003E\u003Cstrong\u003EDate:\u003C\/strong\u003E Thursday, November 19, 2026\u003Cbr\u003E\u003Cstrong\u003ETime:\u003C\/strong\u003E 3:00 p.m. - 4:00 p.m.\u003Cbr\u003E\u003Cstrong\u003ELocation:\u003C\/strong\u003E CSIP Library \u2014 Room 5126, Centergy One, 5th Floor\u003C\/p\u003E\u003Cp\u003E\u003Cstrong\u003EAbstract:\u003C\/strong\u003E Discrete diffusion models achieve strong performance in text and image generation, but inference remains slow\u003Cbr\u003Eand requires balancing sampling efficiency with sample quality. We systematically study how the degree of\u003Cbr\u003Estochasticity in Markov transitions governs this tradeoff, finding that deterministic transitions converge quickly but\u003Cbr\u003Eaccumulate errors, while stochastic transitions converge more slowly but can achieve better final sample quality.\u003Cbr\u003EThrough an information-theoretic analysis, we show that redundant transitions that symmetrically exchange\u003Cbr\u003Eprobability mass between states can contract sampling errors, providing an error-correcting effect. Motivated by\u003Cbr\u003Ethis insight, we introduce Discrete Churn and Restart Sampling (DCRS), which injects controlled stochasticity by\u003Cbr\u003Ealternating between forward and reverse diffusion. Across synthetic, image, and language benchmarks, DCRS\u003Cbr\u003Eimproves the speed-quality tradeoff at low numbers of function evaluations, achieving up to a 10\u00d7 reduction in\u003Cbr\u003Eimage sampling steps while revealing more nuanced behavior for language generation.\u003C\/p\u003E\u003Cp\u003E\u003Cstrong\u003EBio: \u003C\/strong\u003EWilliam Yuan is a Ph.D. student in Electrical and Computer Engineering at the Georgia Institute of Technology,\u003Cbr\u003Eadvised by Prof. Amirali Aghazadeh. His research focuses on developing methods to accelerate the inference in\u003Cbr\u003Edeep generative models. He received his B.S. from Cornell University.\u003C\/p\u003E","summary":"","format":"limited_html"}],"field_subtitle":"","field_summary":[{"value":"\u003Cp\u003EJoin Georgia Tech Ph.D. student William Yuan for a talk on Understanding and Improving Low-Step Discrete Diffusion Sampling.\u003C\/p\u003E","format":"limited_html"}],"field_summary_sentence":[{"value":"Join Georgia Tech Ph.D. student William Yuan for a talk on Understanding and Improving Low-Step Discrete Diffusion Sampling."}],"uid":"36558","created_gmt":"2026-11-03 22:44:36","changed_gmt":"2026-11-03 22:46:28","author":"zwiniecki3","boilerplate_text":"","field_publication":"","field_article_url":"","field_event_time":{"event_time_start":"2026-11-19T15:00:00-05:00","event_time_end":"2026-11-19T16:00:00-05:00","event_time_end_last":"2026-11-19T16:00:00-05:00","gmt_time_start":"2026-11-19 20:00:00","gmt_time_end":"2026-11-19 21:00:00","gmt_time_end_last":"2026-11-19 21:00:00","rrule":null,"timezone":"America\/New_York"},"location":"CSIP Library \u2014 Room 5126, Centergy One, 5th Floor","extras":["free_food"],"groups":[{"id":"1255","name":"School of Electrical and Computer Engineering"}],"categories":[],"keywords":[{"id":"5611","name":"CSIP Seminar Series"}],"core_research_areas":[],"news_room_topics":[],"event_categories":[{"id":"194683","name":"Talk"}],"invited_audience":[{"id":"78761","name":"Faculty\/Staff"},{"id":"177814","name":"Postdoc"},{"id":"174045","name":"Graduate students"},{"id":"78751","name":"Undergraduate students"}],"affiliations":[],"classification":[],"areas_of_expertise":[],"news_and_recent_appearances":[],"phone":[],"contact":[],"email":[],"slides":[],"orientation":[],"userdata":""}}}