New method maps and controls LLM personality traits using OCEAN framework

Researchers have developed a method called "Persona Cartography" to measure and control the personality traits of large language models (LLMs). By adapting the OCEAN framework (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism), they can train low-rank adapters to amplify or suppress specific traits in models ranging from 4B to 32B parameters. These adapters demonstrate a largely monotonic effect on traits as models scale and combine additively, influencing safety-related behaviors like frustration and sycophancy. The approach also includes an unsupervised pipeline to discover interpretable behavioral factors not predefined by human psychometrics. AI

IMPACT Enables more precise control over LLM behavior, potentially improving safety and mitigating undesirable traits like sycophancy.

RANK_REASON The cluster describes a research paper detailing a new methodology for analyzing and manipulating LLM behavior.

Read on arXiv cs.AI →

• agreeableness
• arXiv
• conscientiousness
• extraversion
• Hugging Face
• LLM Judge
• neuroticism
• OCEAN framework
• openness
• Persona Cartography
• Big-5 OCEAN
• Gemma3
• Less Wrong
• Llama 3.1
• Loras
• Ocean
• Open Character Training
• Qwen3

• model release
• paper
• safety

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →