Tag
Introduces fixed-point flows, a self-conditioned flow language model that treats self-conditioning as a fixed-point iteration, enabling distillation into a few-step flow map language model (FMLM⋆) that outperforms prior work on OpenWebText.
This paper reveals that the low generative perplexity (Gen-PPL) reported by continuous diffusion language models like ELF is misleading, as it rewards repetition; the authors identify a one-dimensional attractor in the self-conditioning loop as the cause and propose ACE, a simple fix that subtracts this direction to reduce repetition without sacrificing quality.