Flow Language Models Enhance Reasoning via Continuous Representations

Flow Language Models (FLMs) outperform discrete diffusion models in reasoning efficiency by evolving continuous sequence representations throughout the denoising process. Theoretical analysis suggests that information retained in these continuous states from alternative candidates aids in subsequent solution recovery.