tomg-group-umd

A reasoning model trained to reason in latent space, opposed to the current breed of models which are trained to reason in tokens. The model uses recurrent blocks, which are running multiple times during inference to scale up the inference compute.