Predictive-State Flows for Parallel Language Generation via Diffusion and Decoding
October 12, 2026
Predictive-State Flows (PSF) separates generative roles by using a rectified-flow prior to produce high-level predictive states and a block-causal decoder to resolve token-level choices in parallel. This approach aims to simplify the modeling of joint latent distributions in continuous diffusion language models.
HOW THIS AFFECTS YOU
●
researcherThis architecture provides a new framework for decoupling high-level sequence planning from low-level token decoding in diffusion models.