Persona Perspectives: Adopting Viewpoints Reorganizes Representations in LLMs
arXiv cs.CL · 5 Nov 2026, 04:05 AEST
Research shows that large language models (LLMs) can adopt different personas, affecting their internal representations. This reorganization can lead to politically biased assessments of truthfulness and morality.
Key facts
- 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.
- 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.
- 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.
Read aloud
Research shows that large language models L L M s can adopt different personas, affecting their internal representations. This reorganization can lead to politically biased assessments of truthfulness and morality. Common methods for monitoring features in large language models L L M s such as linear probes and sparse autoencoders SAEs assume that features maintain a consistent representation across contexts. We extract representations of over nine thousand common words and thirty three 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. 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.
Sources in this cluster
- Persona Perspectives: Adopting Viewpoints Reorganizes Representations in LLMs
arXiv cs.CL · 5 Nov 2026, 04:05 AEST