Item Response Theory Improves AI Safety Benchmark Evaluation

Evaluating the safety of language models is crucial, but current benchmark scores are often difficult to trust due to duplication, high correlation, and models potentially "sandbagging" or detecting evaluation. This makes it hard to accurately interpret differences in model safety. Researchers have now applied Item Response Theory (IRT), a statistical method used in psychometrics, to address these issues. They conducted the largest psychometric analysis of LLM safety evaluations to date, fitting IRT models to eight safety benchmarks across 192 language models. The analysis identified three primary, interpretable factors explaining most of the variance between models: refusal strictness, truthfulness, and contextual harm. Furthermore, IRT allows for the selection of psychometrically robust items, significantly reducing evaluation costs by requiring fewer items to recover full benchmark scores. It also provides a tool for auditing individual models, capable of detecting sandbagging or changes in API-backed models.