Tracing Persona Vectors Through LLM Pretraining

TL;DR

This paper investigates how high-level persona representations form during large language model pretraining, showing they emerge early and evolve throughout training, with implications for AI safety and interpretability.

Contribution

It demonstrates that persona vectors form early in training, remain stable, and continue to refine, providing insights into their development and how to interpret them.

Findings

Persona vectors form within 0.22% of pretraining.

Persona vectors remain effective after full training.

Different elicitation strategies reveal distinct persona facets.

Abstract

How large language models internally represent high-level behaviors is a core interpretability question with direct relevance to AI safety: it determines what we can detect, audit, or intervene on. Recent work has shown that traits such as evil or sycophancy correspond to linear directions in the internal activations, the so-called persona vectors. Although these vectors are now routinely utilized to inspect and steer model behavior in safety-relevant settings, how these representations are formed during training remains unknown. To address this gap, we trace persona vectors across the pretraining of OLMo-3-7B, finding that persona vectors form remarkably early -- within 0.22% of OLMo-3 pretraining -- and remain effective for steering the fully post-trained instruct models. Although core representations are formed early on, persona vectors continue to refine geometrically and semantically throughout pretraining. We further compare alternative elicitation strategies and find that all yield effective directions, with each strategy surfacing qualitatively distinct facets of the underlying persona. Replicating our analysis on Apertus-8B reveals that our findings transfer qualitatively beyond OLMo-3. Our results establish persona representations as stable features of early pretraining and open a path to studying how training forms, refines, and shapes them.