Tracing Persona Vectors Through LLM Pretraining

Persona traits form extremely early in AI training: The study found that behavioral traits like "evil" or "sycophancy" become detectable within just 0.22% of the pretraining process for OLMo-3-7B, suggesting these problematic behaviors emerge almost immediately rather than developing gradually over time.

Early-formed traits remain controllable throughout training: Once persona vectors form early in pretraining, they remain effective tools for steering model behavior even in the final, fully-trained models, indicating that early interventions could potentially have lasting safety benefits.

Multiple ways to detect the same underlying behaviors: The researchers found that different methods of eliciting persona traits all surface effective directions in the model's internal representations, though each method reveals different facets of the same underlying behavior—suggesting safety audits may need multiple approaches to fully understand model tendencies.

These findings generalize across different models: The results held up when tested on a second model (Apertus-8B), indicating that persona vector formation during early pretraining is likely a fundamental feature of how large language models develop, not specific to one architecture.

Internal representations are more stable and interpretable than previously understood: By showing that persona vectors maintain geometric and semantic consistency from early training onward, this work suggests AI safety researchers have more reliable tools for detecting and potentially intervening on problematic behaviors than previously thought.