The authors start from the observation that linear probes trained to detect deception as one direction in representation space fail to generalize across different deception scenarios.
Deception in LLMs is not a single direction, adaptive subspace probes work better
Dynamic Deception Subspace Navigation with 10-shot gating reaches 0.97 AUROC across datasets, beating static linear probes.
Top University
Yao Huang · Siyuan Chen · Zhenghan Zhou · Yinpeng Dong
Tsinghua University · Shanghai Qi Zhi Institute
Research Digest··2 min read
The authors show that deceptive intent in large language models does not align with a single universal direction in activation space.
§