Parallelizable Probabilistic Latent Reasoning for Diverse Trajectories

To decouple reasoning from linguistic token generation, this method moves Chain-of-Thought processes into a compact, continuous latent space. By modeling a distribution over latent steps rather than single paths, it enables parallel exploration of diverse reasoning trajectories without the overhead of autoregressive verbalization.