Tag
This paper systematically studies how lie typology, representation depth, probe expressivity, and sparse features impact deception detection in LLMs, finding that detection performance is highly dependent on training data and representation choice.