Linear encoding of features with liar counterexamples

Features are encoded as linear directions in activation space, and the model performs linear operations over them (e.g. antonym = reflection, negation = negation). The authors also identify "liar" counterexamples where a linear probe disagrees with the model's actual behavior, and show the geometry of activation space is isomorphic to the logical structure of the underlying concepts.