To Make Language Models Work Better, Researchers Sidestep Language

I wonder why they go with recurrent rather than something like latent flow-matching?