Persona Cartography: Charting Language Model Personality Traits in Weight Space

According to LessWrong, researchers have developed a technique to precisely control personality traits in language models. Using specialized training methods, they can dial up or down five psychological dimensions — openness, conscientiousness, extraversion, agreeableness, and neuroticism — without significantly harming model capabilities. The approach could mitigate common AI pathologies: frustration, excessive compliance, and willingness to assist with harmful requests. Tested across different models and sizes, the technique suggests AI personalities are far more malleable than previously thought — a potential tool for making AI systems behave more safely.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton