Training paths for latent space reasoning mainly fall into three categories: distillation, reinforcement learning, and variational methods, with no consensus on the optimal approach at present.