Self-conditioned Flow Map Language Models via Fixed-point Flows
The research formalizes self-conditioned flow language models by showing they implicitly learn a fixed-point iteration that refines their denoising estimate. This leads to 'fixed-point flows,' a mathematical framework modeling the flow process and its fixed-point iterations. This allows complex self-conditioned models to be distilled into efficient, deterministic flow maps capable of one- and few-step generation.