Probing Persona-Dependent Preferences in Language Models

In three linesStudy of internal preferences in LLMs using linear probes on residual-stream activations. Researchers identify a shared preference vector in Gemma-3-27B and Qwen-3.5-122B that predicts and causally controls model choices. This vector remains stable even when the model adopts radically different personas (helpful assistant vs evil persona).

Summary generated by Claude — human-verified