Linear Probes May Generalize Better Using Persona Coordinates for AI Safety

A new arXiv preprint (2605.09391) investigates whether linear probes, a white-box monitoring method, can generalize better under distribution shift by operating in a low-dimensional subspace of model internals. The authors propose constructing persona axes for deception and sycophancy using contrastive persona prompts, inspired by the Assistant Axis and Persona Selection Model. Unsupervised PCA of persona-specific vectors yields first principal components that cleanly separate harmful behaviors, potentially improving robustness against strategic deception and sandbagging in language models. The study addresses the failure of current probes under distribution shift, which limits their real-world utility for monitoring harmful behaviors during model interactions.