On Reasoning Correctness Directions.

arXiv:2610.25464v1 Announce Type: new Abstract: We present experimental evidence suggesting that information about whether a model's reasoning is on the right track can be linearly decodable without defining a causally useful direction for steering. To study t