Latent Chain-of-Thought as Planning: Decoupling Reasoning from Verbalization
PLaT reformulates complex reasoning by treating it as planning within a continuous latent space, separating thought generation from text output. It uses a Planner to evolve latent states and a Decoder to verbalize them, allowing for dynamic stopping of reasoning. This method offers better diversity and scalability than traditional Chain-of-Thought methods while being significantly faster.