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

Researchers have discovered that an AI model's internal processing states correspond to specific types of reasoning, such as deduction or mathematical calculation. These distinct patterns emerge primarily within the middle layers of the neural network during task execution. This finding indicates that models perform underlying computations that are not explicitly captured in their visible chain-of-thought outputs. Understanding these hidden processes could improve AI safety by providing better insight into how models arrive at their conclusions.