Do Linear Probes Generalize Better in Persona Coordinates?

• URL: http://arxiv.org/abs/2605.09391v2
• Date: Fri, 15 May 2026 12:37:57 GMT
• Title: Do Linear Probes Generalize Better in Persona Coordinates?
• Abstract summary: We study whether there exists a low-dimensional subspace of the model internals that captures harmful behaviors more robustly.<n>We construct persona axes for deception and sycophancy using contrastive persona prompts.<n>We show that persona-derived directions transfer non-trivially and probes trained on persona-PC projections generalize better than probes trained on raw activations.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• 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.

Related papers

• World Models for Embodied Intelligence: From Plausible to Controllable to Actionable [100.13060753774657]
  World models connect perception and decision-making in embodied intelligence by maintaining hidden state, anticipating consequences, comparing interventions, and adapting when execution departs from expectations.<n>This raises a central question: which predictive capabilities improve behavior?<n>We introduce three progressively stronger capability levels: Plausible models preserve task-relevant temporal, geometric, or physical structure; Controllable models additionally predict how interventions alter that structure; and Actionable models translate predictions into measurable gains in planning, action, learning, evaluation, verification, recovery, or data selection.
  arXiv  Detail & Related papers  (2026-09-15T06:22:42Z)
• Revealing Hidden Model Behaviors with Task-Specific Self-Reports [0.0]
  We introduce the Stabilized Adapter for self-Report (SAR), a lightweight LoRA adapter that makes a fine-tuned model describe its own hidden behavior in plain language.<n>Across seven implanted behaviors (plus a no-behavior control), SAR detects the hidden behavior in every one--even when the model has generalized into broad misalignment.
  arXiv  Detail & Related papers  (2026-07-03T23:41:25Z)
• The Illusion of Intervention: Your LLM-Simulated Experiment is an Observational Study [56.649987197786096]
  Large language models (LLMs) show potential as simulators of human behavior.<n>LLMs can induce unintended shifts in latent user attributes, causing user drift.<n>We formalize the confounding or selection bias that can arise due to user drift.
  arXiv  Detail & Related papers  (2026-05-20T06:09:41Z)
• Tracing Persona Vectors Through LLM Pretraining [16.7064056526422]
  How large language models internally represent high-level behaviors is a core interpretability question with direct relevance to AI safety.<n>Recent work has shown that traits such as evil or sycophancy correspond to linear directions in the internal activations, the so-called persona vectors.<n>We find that persona vectors form remarkably early -- within 0.22% of OLMo-3 pretraining -- and remain effective for steering the fully post-trained instruct models.
  arXiv  Detail & Related papers  (2026-05-13T10:44:23Z)
• Building Better Deception Probes Using Targeted Instruction Pairs [1.610762469264735]
  Linear probes are a promising approach for monitoring AI systems for deceptive behaviour.<n>In this paper, we identify the importance of the instruction pair used during training.<n>We show that targeting specific deceptive behaviors through a human-interpretable taxonomy of deception leads to improved results on evaluation datasets.
  arXiv  Detail & Related papers  (2026-02-01T20:18:11Z)
• PersonaDrift: A Benchmark for Temporal Anomaly Detection in Language-Based Dementia Monitoring [0.9668407688201359]
  PersonaDrift is a benchmark designed to evaluate machine learning and statistical methods for detecting progressive changes in daily communication.<n>The benchmark focuses on two forms of longitudinal change that caregivers highlighted as particularly salient: flattened sentiment and off-topic replies.<n>Preliminary results show that flattened sentiment can often be detected with simple statistical models in users with low baseline variability.
  arXiv  Detail & Related papers  (2025-11-20T15:15:00Z)
• Impatient Users Confuse AI Agents: High-fidelity Simulations of Human Traits for Testing Agents [58.00130492861884]
  TraitBasis is a lightweight, model-agnostic method for systematically stress testing AI agents.<n>TraitBasis learns directions in activation space corresponding to steerable user traits.<n>We observe on average a 2%-30% performance degradation on $tau$-Trait across frontier models.
  arXiv  Detail & Related papers  (2025-10-06T05:03:57Z)
• Persona Vectors: Monitoring and Controlling Character Traits in Language Models [11.039979968884575]
  Large language models interact with users through a simulated 'Assistant' persona.<n>We identify directions in the model's activation space-persona vectors.<n>We apply persona vectors to predict and control personality shifts that occur during training.
  arXiv  Detail & Related papers  (2025-07-29T05:20:14Z)
• Modeling User Preferences via Brain-Computer Interfacing [54.3727087164445]
  We use Brain-Computer Interfacing technology to infer users' preferences, their attentional correlates towards visual content, and their associations with affective experience. We link these to relevant applications, such as information retrieval, personalized steering of generative models, and crowdsourcing population estimates of affective experiences.
  arXiv  Detail & Related papers  (2024-05-15T20:41:46Z)
• Pedestrian Detection: Domain Generalization, CNNs, Transformers and Beyond [82.37430109152383]
  We show that, current pedestrian detectors poorly handle even small domain shifts in cross-dataset evaluation. We attribute the limited generalization to two main factors, the method and the current sources of data. We propose a progressive fine-tuning strategy which improves generalization.
  arXiv  Detail & Related papers  (2022-01-10T06:00:26Z)
• Tracking the risk of a deployed model and detecting harmful distribution shifts [105.27463615756733]
  In practice, it may make sense to ignore benign shifts, under which the performance of a deployed model does not degrade substantially. We argue that a sensible method for firing off a warning has to both (a) detect harmful shifts while ignoring benign ones, and (b) allow continuous monitoring of model performance without increasing the false alarm rate.
  arXiv  Detail & Related papers  (2021-10-12T17:21:41Z)
• DAAIN: Detection of Anomalous and Adversarial Input using Normalizing Flows [52.31831255787147]
  We introduce a novel technique, DAAIN, to detect out-of-distribution (OOD) inputs and adversarial attacks (AA) Our approach monitors the inner workings of a neural network and learns a density estimator of the activation distribution. Our model can be trained on a single GPU making it compute efficient and deployable without requiring specialized accelerators.
  arXiv  Detail & Related papers  (2021-05-30T22:07:13Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.