Continuous Latent Reasoning

Continuous latent reasoning asks whether language and multimodal models can solve problems by updating internal representations rather than generating a word-by-word chain of thought, potentially reducing latency and token use. Approaches include recurrent or looped latent rollouts, learned latent prefixes, flow- and diffusion-based trajectories, reinforcement learning, and adaptive rules for how long to think. These methods are tested on arithmetic, Sudoku and maze planning, code generation, visual reasoning, audio tasks, and tool-using agents, with reported gains in efficiency and sometimes accuracy over explicit reasoning. A central open tension is whether compact latent states faithfully support the answer: accuracy alone does not establish this, and causal analyses find shortcut behavior while robustness, supervision, and reliable decoding remain challenges.