Persona prompting reorganizes LLM latent representations
Original title: Persona Perspectives: Adopting Viewpoints Reorganizes Representations in LLMs
LLM persona adoption alters semantic axis projections in ideological directions.
Common feature extraction methods assume consistent representations in LLMs. However, prompting models to adopt liberal or conservative personas changes the geometry of latent space. This affects probes for truthfulness and morality, leading to slanted evaluations on controversial statements.
- Over 9,000 words and 33 semantic axes analyzed
- Persona conditioning systematically alters projections and alignments
- Probes produce politically slanted ratings of truth and morality
- AI models
- LLM
From the original article
1 min read at arxiv.orgCommon 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.
By Austin C. Kozlowski · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.