Probing persona-dependent preferences in language models

Oscar Gilg, Pierre Beckmann, Daniel Paleka, Patrick Butlin · 2026 · arXiv

Read the paper

Finds a preference direction in two open models that predicts and steers task choices. The same direction serves very different personas, including an ‘evil’ one whose preferences run opposite to the Assistant’s, which suggests shared preference machinery read out by whichever persona is active.

Questions it bears on

• Whose valence is it
• Coherence across personas
• Do models represent goals and preferences