LLM self-reports showed selective links to behavior in tests of 11 models

A researcher says tests of 11 frontier LLMs across four behavioral tasks found selective links between psychometric self-reports and behavior. Within a shared conversation, the Theory of Planned Behavior reached human-level coherence; broad Big Five traits did not. Across separate conversations, coherence held for behaviors such as implicit bias but collapsed for context-driven behaviors such as sycophancy. Persona prompting made self-reports more consistent without bringing behavior into alignment.