NF-CoT
Latent reasoning with normalizing flows — continuous chain-of-thought
A latent-reasoning framework that performs intermediate computation in a continuous space using normalizing flows, while preserving the advantages that make chain-of-thought effective in autoregressive LLMs: native left-to-right generation, probabilistic sampling, KV-cache-compatible decoding, and tractable likelihoods. Joint work with Prof. Jiatao Gu and the GMLR group at the University of Pennsylvania. Accepted to the COLM 2026 Workshop on Efficient Reasoning.