Reasoning steps like calculation, formula retrieval, and deduction are clearly separable in a model's internal states, especially in the middle layers. That matters for AI safety, because models process more than their visible chain of thought reveals. The article AI models' written reasoning steps
