Beyond Next Token Prediction: Diffusion and Flow Models for Next-Generation Decoding

Researchers explore discrete diffusion and flow models as alternatives to sequential next-token generation. The workshop spans mathematical foundations, efficient algorithms, scalable implementations and evaluation across language and scientific data, considering both the capabilities and limitations of parallel decoding. Session: 08:00–17:00 (Australia/Sydney). A paid NeurIPS virtual pass includes workshop livestreams; on-site access follows the host-location registration rules.