Chanhyuk Lee

Looped flows are proposed, an approach that allows solving harder problems by spending more computation through a finer temporal grid and enables multiple valid predictions from different initial noise samples.

Fixed-point flows enable a leap in performance for language models, outperforming state-of-the-art methods in one- and few-step generation tasks.

Achieving high-fidelity language generation with 32x fewer function evaluations could revolutionize real-time applications of language models.

Forget slow, multi-step diffusion: this work achieves state-of-the-art text generation quality with a *single* denoising step using flow-based language models.