Looped Transformers & Latent Reasoning

This direction tracks architectures and methods that repeat computation along the depth axis, covering looped and recurrent-depth transformers with weight-shared layer iteration, recursive reasoning models such as HRM and TRM, latent reasoning carried out in depth, and adaptive computation that allocates depth per token through early exiting, dynamic layer skipping, and mixture-of-depths. The central question is whether loop count can become a scalable compute axis alongside parameter count, with emphasis on comparisons against parameter- and compute-matched non-looped models and on empirical evidence of extrapolation in the number of iterations.