🤖 AI Summary
This work addresses the limited generalization of existing linear probes based on internal model activations under distribution shifts, which hinders reliable monitoring of harmful behaviors in language models. To overcome this, the authors propose incorporating personality coordinates as an inductive bias: they construct a deception–ingratiation personality axis using contrastive personality prompts and apply unsupervised PCA to extract dominant directions from activation spaces. Linear probes are then constrained to this low-dimensional subspace during training. This approach substantially enhances probe robustness and transferability across datasets and behavioral contexts. Experiments on ten evaluation benchmarks demonstrate that probes trained on personality-based principal component projections consistently outperform baselines trained directly on raw activations, with a unified personality axis further improving generalization performance.
📝 Abstract
It is becoming increasingly necessary to have monitors check for harmful behaviors during language model interactions, but text-only monitoring has not been sufficient. This is because models sometimes exhibit strategic deception and sandbagging, changing their behavior during evaluation. This motivates the use of white-box monitors like linear probes, which can read the model internals directly. Currently, such probes can fail under distribution shift, limiting their usefulness in real settings. We study whether there exists a low-dimensional subspace of the model internals that captures harmful behaviors more robustly, while leaving out spuriously correlative features. Inspired by the Assistant Axis and Persona Selection Model, we construct persona axes for deception and sycophancy using contrastive persona prompts. The first principal components, obtained by unsupervised PCA of the persona-specific vectors, cleanly separate harmful and harmless personas. Across 10 evaluation datasets, we show that persona-derived directions transfer non-trivially and probes trained on persona-PC projections generalize better than probes trained on raw activations. We also find that a unified axis consisting of multiple harmful and harmless behaviors improves generalization across behaviors and datasets. Overall, persona vectors provide a useful inductive bias for building more transferable behavior probes.