AI models' written reasoning steps correspond to distinct internal patterns, a new study finds

A new study reveals that reasoning steps in AI models, such as calculations and formula retrieval, can be distinctly separated in the model's internal states. The findings indicate that models process more information than their visible chain of thought suggests, highlighting the importance of understanding these patterns in the context of AI safety. This discovery could influence how AI models are developed, as developers need to consider these internal patterns to ensure the safety and reliability of systems.