The Unsampled Truth: Quantifying Prompt Artifacts in LM Psychometrics
When prompting language models for psychometric assessment, researchers assume that the responses reflect the injected persona and the meaning of the survey item. We test this premise using a diagnostic design that crosses five semantically distinct baseline personas with five semantically equivalent variants of each o...
Nils Schwager, Christoph Hau, Simon Münker et al.
· 0 citations