Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning: This paper is an interesting take at expanding the reasoning expressiveness of models. Instead of having the models reason in latent space with non-discrete intermediate outputs during inference (a la COCONUT), they train the model to learn a codebook of discrete latent tokens during training and then use these tokens during inference. By specifying a threshold difference between successive states in the recurrent portion, the model can keep iterating until it has converged on a token – allowing for more iterations on “difficult” tokens and fewer on “easier” tokens that converge more quickly.