BenchMIRT: What are LLM benchmarks actually measuring?

LLM benchmarks often measure more than their intended ability, according to a new method called BenchMIRT. Researchers have developed a way to dissect the individual prompts within a benchmark to reveal what specific skills they are actually assessing. By applying Item Response Theory at both the model and question levels, BenchMIRT can separate multiple capabilities that contribute to a model's performance.

This technique has been trained on benchmarking results from 100 LLMs across 16 benchmarks, revealing two dominant dimensions: safety and general reasoning. Some benchmarks, like BBQ which tests social stereotypes, align more with general reasoning than safety. Others, like WMDP which evaluates dangerous knowledge, show a stronger connection to general reasoning.

HarmBench, which tests compliance with harmful requests, demonstrates how a single benchmark can incorporate various signals, with both safety and general reasoning dimensions strongly associated. Overall, BenchMIRT highlights that a single benchmark score can combine multiple signals, offering a more nuanced interpretation of a model's abilities.