Persona Cartography — mapping LLM personality in weight space
Researchers trained Big-5 OCEAN personality traits into LLMs (Llama, Qwen, Gemma; 4B–32B) via modified LoRA adapters, then manipulated them with weight arithmetic.
• Scale, invert, and combine trait LoRAs to amplify or suppress behaviours
• Mitigate common LLM pathologies by blending personas
• Unsupervised discovery of emergent trait patterns not pre-defined
Implications: personality traits live in predictable regions of weight space and can be treated as composable, orthogonal components for controllable alignment.