A research team has built a system for measuring, editing, and combining personality traits directly inside language model weights.
The paper, posted to arXiv, treats model behavior as positions in a "personality space" defined by the OCEAN framework — Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. The team trained low-rank adapters, small plug-in weight modifications, to amplify or suppress each trait independently across six models ranging from 4B to 32B parameters. They validated the results using an LLM-based judge calibrated against human raters, trait-specific benchmarks, and standard capability tests. The adapters also compose: stacking multiple trait adapters produces roughly additive effects, letting researchers dial up a "conscientious but low-agreeableness" model without retraining from scratch.
The safety implications are the part worth paying attention to. The researchers found that nudging a model along the neuroticism axis changed how it expressed frustration, while the agreeableness axis directly affected sycophancy — the well-documented tendency of models to tell users what they want to hear. That gives safety teams a concrete, measurable handle on behaviors that have mostly been managed through prompt engineering and RLHF guesswork.
Personality control in AI is not a new idea — character prompting has been a staple of chatbot products for years — but baking traits into weights rather than system prompts means the behavior is harder to talk the model out of, which is either a feature or a risk depending on who is deploying it.