LLM Persona Encoding Found in Final Decoder Layers

A new study from arXiv (2505.24539) investigates how personas—defined by human characteristics, values, and beliefs—are encoded in large language models. Using dimension reduction and pattern recognition, researchers identified the final third of decoder layers as the region where persona representations diverge most. Within these layers, overlapping activations were observed for ethical perspectives like moral nihilism and utilitarianism, indicating polysemy, while political ideologies such as conservatism and liberalism showed distinct patterns. The findings span multiple pre-trained decoder-only LLMs.