Tracing Psychometric Inference in Large Language Models
Summary
- Researchers investigated whether large language models internally represent psychological constructs or merely mimic psychometric structures in their outputs.
- Using a cross-persona paradigm across 14 variants of Llama 3 and Qwen 2.5, models predicted responses across seven psychological scales based on data from 272 university students.
- Larger instruct variants displayed internal geometry aligned with human correlation matrices and behavioral amplification.
- Activation-probing revealed a probe-behavior gap termed readout attenuation, demonstrating decodability without proving mediation.
Why it matters
This research provides insights into how large language models internally encode psychological constructs and human behavioral traits, helping researchers understand the mechanisms behind persona-based prompting and psychological inference.