LLM-NRM Psychometric Framework for Multiple-Choice Benchmarking

The LLM Nominal Response Model (LLM-NRM) uses option-level psychometrics to evaluate LLM performance beyond binary correct/incorrect scores. By modeling the full distribution of answer choices, it achieves a 0.920 Spearman correlation with held-out interactions, capturing model calibration and positional biases.