This research introduces a novel approach for assessing LLM personality traits by simulating multiple independent observers rather than relying on potentially biased self-reports.

• Addresses limitations of traditional self-report questionnaires which fail to capture true behavioral nuances
• Draws inspiration from informant-report methods in human psychology
• Creates a more robust evaluation framework that reduces meta-knowledge contamination
• Enables more accurate detection of potential biases and security vulnerabilities in LLM behavior

Security Implications: Understanding LLM personality traits through multiple perspectives helps identify behavioral patterns that could lead to misuse, providing a stronger foundation for safer AI deployment in sensitive contexts.

Beyond Self-Reports: Multi-Observer Agents for Personality Assessment in Large Language Models