Persona Perspectives: Adopting Viewpoints Reorganizes Representations in LLMs
Political persona prompts changed LLM representation geometry, causing probes trained without personas to give divergent, politically slanted truthfulness and morality ratings.
The study compares representations for **over 9,000 words** across **33 semantic axes**, using liberal and conservative personas and two neutral baselines. Persona conditioning changed both word projections and relationships among axes; classifiers trained without personas then produced divergent truth and morality ratings on controversial statements.
If you use probes or sparse autoencoders to monitor model behavior, check them under the same persona or prompt conditions as the system being evaluated. A neutral-context reading may not carry over to a prompted viewpoint.
The study compares representations for **over 9,000 words** across **33 semantic axes**, using liberal and conservative personas and two neutral baselines. Persona conditioning changed both word projections and relationships among axes; classifiers trained without personas then produced divergent truth and morality ratings on controversial statements. If you use probes or sparse autoencoders to monitor model behavior, check them under the same persona or prompt conditions as the system being evaluated. A neutral-context reading may not carry over to a prompted viewpoint. The reported tests focus on political personas and truthfulness and morality features. The material does not establish how broadly the effects extend to other personas or monitoring tasks.
This adds evidence that persona prompts can change the model’s internal semantic organization, not just its outward answers: neutral-context probes may misread truth and morality features under a prompted viewpoint. It supports evaluating monitors under the same persona conditions as the system, while the political-persona scope leaves broader generalization open.