Study Maps Where Personas Are Encoded Inside Large Language Models

Researchers have conducted a systematic investigation into the internal architecture of large language models to understand where and how personas—sets of human characteristics, values, and beliefs—get encoded in their representation space.

How the Research Was Conducted#

The team applied dimensionality reduction and pattern recognition techniques to trace which layers in multiple pre-trained decoder-only LLMs encode persona information most distinctly. They then examined the neural activations within specific layers to compare how different personas are positioned relative to one another, mapping out their shared and separate embedding territories.

Key Findings on Persona Encoding#

The analysis revealed that persona representations diverge most significantly in the final third of decoder layers. The research uncovered an important distinction in how different types of perspectives are organized:

• Ethical viewpoints like moral nihilism and utilitarianism show overlapping activations, suggesting the model encodes them as semantically related concepts
• Political ideologies such as conservatism and liberalism occupy more separated regions of representation space, indicating they are treated as more categorically distinct

This pattern of partial overlap—known as polysemy—reveals how LLMs internally conflate certain philosophical positions while keeping others architecturally separate.

Implications for Model Control#

These findings offer a foundation for future work aimed at steering or refining how LLMs express specific human traits in their outputs. By identifying where and how personas live in the model's neural space, researchers can potentially develop more targeted approaches to controlling which viewpoints emerge during generation.