Probe Generalization as Subspace Selection for OOD Deception Detection

An AI deception detector can learn patterns peculiar to its training examples and then struggle in a different setting. This study tests detectors that read a language model's internal activity. Researchers separated that activity into statistical patterns and used another language model to identify patterns whose descriptions suggested broader relevance to deception. Restricting the detector to those selected patterns improved transfer across three test datasets, although one offered little room for improvement. The result suggests that choosing which internal signals a detector uses matters for transfer. The experiments cover one underlying model and one training dataset. Methods and limitations.