Found 1 document(s)
Page 1 of 1
-
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 this, we ...