Do Linear Probes Generalize Better in Persona Coordinates?
On the map: Mechanistic interpretability → Activation steering methods
Abstract (arXiv)
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.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Activation steering methods, one of 13 directions in Mechanistic interpretability. Builds on 16 of its 32 references that are in the map. First posted on arXiv 10 May 2026.
Builds on
16 of 32 references.
Closest ancestors · 1
Same direction, newest first
Other references · 15
- 2016Understanding intermediate layers using linear classifier probescited by 1,353
- 2023Towards Understanding Sycophancy in Language Modelscited by 1,040
- 2022Discovering Latent Knowledge in Language Models Without Supervisioncited by 767
- 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
- 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 712
- 2023The Internal State of an LLM Knows When It's Lyingcited by 659
- 2022Language Models as Agent Modelscited by 132
- 2026The Assistant Axis: Situating and Stabilizing the Default Persona of Language Modelscited by 90
- 2023How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questionscited by 83
- 2025Detecting Strategic Deception Using Linear Probescited by 68
- 2023Personas as a Way to Model Truthfulness in Language Modelscited by 47
- 2026Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreakscited by 41
3 more
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors
- 2024Who's asking? User personas and the mechanics of latent misalignment
- 2026Extracting Persona Subspaces Through Iterative Nullspace Projection For Modulation
- 2026Persona Prompting as a Lens on LLM Social Reasoning
- 2026Characterizing the Consistency of the Emergent Misalignment Persona
- 2025Evaluating Chinese Large Language Models: The Influence of Persona Assignment on Stereotypes and Safeguards
- 2026The Persona Hierarchy Model: Understanding Contextual Generalization in Fine-Tuning LLMs
- 2026Persona-Model Collapse in Emergent Misalignment