Researchers report experimental evidence that information about whether a language model's reasoning is on the right track can be linearly decoded from its internal representations. In other words, a simple linear readout can detect whether the model is heading toward a correct answer.
However, the same information does not appear to define a direction that is causally useful for steering the model's behavior. The finding separates the ability to observe reasoning quality from the ability to intervene on it, suggesting that knowing a model is wrong is not the same as having a lever to fix it.
The authors say they introduce a new approach to study this question, but the abstract as provided cuts off before describing it. The paper is available on arXiv under identifier 2610.25464.