NF-CoT: Latent reasoning with normalizing flows

Category: ML research
Status: developing
Sources: arXiv:2606.06447, project page
Summary: NF-CoT proposes performing intermediate reasoning steps in compact continuous states using normalizing flows before committing to text, rather than forcing computation through a discrete token stream. The approach preserves left-to-right generation, probabilistic sampling, KV-cache compatibility, and tractable likelihood estimation while reducing the verbalization overhead of standard chain-of-thought.
Why it matters: If the continuous reasoning overhead proves lower than discrete CoT at equivalent accuracy, inference costs for reasoning-heavy tasks could decrease without requiring separate latent-space model architectures.