aiqbench analysis
This research demonstrates that prompting LLMs to adopt political personas systematically reorganizes their latent space geometry, causing standard monitoring tools like linear probes and SAEs to produce ideologically biased assessments of truthfulness and morality. The findings highlight a critical vulnerability in current AI safety protocols, as features extracted under neutral conditions fail to accurately reflect the model's behavior when conditioned on specific viewpoints. This underscores the need for context-aware evaluation methods that account for dynamic representational shifts induced by persona adoption.
Persona Perspectives: Adopting Viewpoints Reorganizes Representations in LLMs
arXiv:2611.06848v1 Announce Type: new Abstract: Common methods for monitoring features in large language models (LLMs) such as linear probes and sparse autoencoders (SAEs) assume that features maintain a consistent representation across contexts. LLMs, however, can flexibly adopt different perspectives or "personas" that suggest different underlying systems of conceptual relations. In this paper, we examine whether prompting an LLM to adopt a persona alters the representational geometry of the model's latent space. We extract representations of over 9,000 common words and 33 semantic axes from models adopting liberal and conservative political personas and compare these against two neutral baselines. We find that persona conditioning systematically alters both the projections of words onto semantic axes and the alignments among the axes themselves in directions consistent with the specified ideology. We then test the implications of this reorganization for two safety-relevant features: truthfulness and morality. Linear probes, SAE-based classifiers, and classifiers built from manually selected word representations, all trained without a persona, produce divergent results when applied to models conditioned on political personas, producing politically slanted ratings of truth and morality on controversial statements. These findings suggest that internal representations, as they are conventionally extracted, can exhibit different geometric relations when the model adopts personas, potentially undermining monitoring efforts.
Abstract:Common methods for monitoring features in large language models (LLMs) such as linear probes and sparse autoencoders (SAEs) assume that features maintain a consistent representation across contexts. LLMs, however, can flexibly adopt different perspectives or "personas" that suggest different underlying systems of conceptual relations. In this paper, we examine whether prompting an LLM to adopt a persona alters the representational geometry of the model's latent space. We extract representations of over 9,000 common words and 33 semantic axes from models adopting liberal and conservative political personas and compare these against two neutral baselines. We find that persona conditioning systematically alters both the projections of words onto semantic axes and the alignments among the axes themselves in directions consistent with the specified ideology. We then test the implications of this reorganization for two safety-relevant features: truthfulness and morality. Linear probes, SAE-based classifiers, and classifiers built from manually selected word representations, all trained without a persona, produce divergent results when applied to models conditioned on political personas, producing politically slanted ratings of truth and morality on controversial statements. These findings suggest that internal representations, as they are conventionally extracted, can exhibit different geometric relations when the model adopts personas, potentially undermining monitoring efforts.
Computation and Language (cs.CL)
arXiv:2611.06848 [cs.CL]
(or arXiv:2611.06848v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2611.06848 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 296 words
Related coverage
Lens: Bringing the Right Semantic Perspective into Focus for Training-Free Multimodal Representation Learning
arXiv · Sep 18, 2026
Order-Invariant Answers, Order-Sensitive Representations in Mathematical Reasoning
arXiv · Sep 24, 2026
Mitigating Bias in Large Vision-Language Models via Counterfactual Ensemble Decoding
arXiv · Aug 25, 2026
Complementary Supervised and Self-Supervised Representations for Out-of-Distribution Graph Learning
arXiv · Oct 7, 2026