Persona Cartography maps LLM personality traits with weight arithmetic

Persona Cartography introduces a method to instill and control personality traits in large language models using low-rank adapters (LoRAs). The team trained LoRAs for each of the Big-5 OCEAN traits (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) on multiple model families (Llama 3.1, Qwen3, Gemma3) across sizes from 4B to 32B parameters. Their pipeline modifies the Open Character Training approach: they use a constitution to generate DPO pairs and then deeply instill traits via self-reflection and self-interaction rollouts.

The key innovation is that these trait-specific LoRAs can be manipulated with simple weight-matrix arithmetic. Users can scale a trait up or down, invert it (e.g., turn agreeableness into disagreeableness), or combine multiple traits by adding their corresponding LoRAs. This enables fine-grained, inference-time control over the model's persona without expensive full fine-tuning. The researchers demonstrate that this approach can mitigate common LLM pathologies like frustration and sycophancy, and can reduce willingness to assist with dangerous requests.

Beyond predefined psychometric traits, Persona Cartography also proposes an unsupervised method to discover latent behavioral traits that emerge naturally in the model's weight space. This is important because LLMs may develop weird personas that don't map neatly onto human psychology. The work positions persona as a critical lever for AI alignment: dispositions are upstream of instrumental convergence, and controlling character could help ensure models are safe by default.