Evaluation Drift in LLM Personality Induction: Are We Moving the Goalpost?

In three linesStudy on personality induction in LLMs via fine-tuning (SFT, DPO, ORPO) on long-form essays linked to Big Five profiles. Fine-tuning reduces variance in IPIP-NEO questionnaire responses, but accuracy on full personality profiles remains near chance. Unguided essays lack sufficient cues for faithful personality expression.

Summary generated by Claude — human-verified