Probing Persona-Dependent Preferences in Language Models
For AI alignment researchers, this reveals a mechanistic basis for preference generalization across personas, though the finding is limited to two large models and specific probe methods.
The authors identify a shared neural representation of preferences in LLMs that generalizes across personas, including anti-correlated ones, and demonstrate causal control via steering on Gemma-3-27B.
Large language models (LLMs) can be said to have preferences: they reliably pick certain tasks and outputs over others, and preferences shaped by post-training and system prompts appear to shape much of their behaviour. But models can also adopt different personas which have radically different preferences. How is this implemented internally? Does each persona run on its own preference machinery, or is something shared underneath? We train linear probes on residual-stream activations of Gemma-3-27B and Qwen-3.5-122B to predict revealed pairwise task choices, and identify a genuine preference vector: it tracks the model's preferences as they shift across a range of prompts and situations, and on Gemma-3-27B steering along it causally controls pairwise choice. This preference representation is largely shared across personas: a probe trained on the helpful assistant predicts and steers the choices of qualitatively different personas, including an evil persona whose preferences anti-correlate with those of the Assistant.