Latent Reasoning
Overview
Latent reasoning asks whether models can solve problems by updating internal representations rather than generating every intermediate step as text, potentially reducing inference cost while preserving useful computation. Approaches include recurrent and looped transformers, continuous or discrete latent thought tokens, learned memory and refinement, and hybrid systems that switch between hidden computation and explicit reasoning; training uses curricula, answer or process supervision, and reinforcement learning. Methods are tested on tasks such as arithmetic and competition math, multi-hop question answering, structured search, coding, and multimodal reasoning, with some matching explicit chain-of-thought performance using shorter computations. A central open tension is whether these gains reflect faithful, generalizable reasoning rather than shortcuts: accuracy alone is insufficient, so causal interventions, matched controls, and tests on harder or longer problems are important.
Latest papers
74 papers6 papers in the last four weeks, down 50% on the four weeks before. 0.0% of all new papers.