BenchMIRT: What are LLM benchmarks actually measuring?
Ai2 has introduced BenchMIRT, a novel auditing tool designed to examine what large language model benchmarks evaluate at the item level. By adapting multidimensional item response theory from psychometrics, the framework analyzes prompt performance to uncover underlying model abilities. The system was tested on evaluation data involving 100 LLMs, 16 benchmarks, and over 34K questions.
Key Takeaways
- Researchers at Ai2 created BenchMIRT to examine the specific capabilities driving performance scores in AI evaluations.
Traditional benchmarking often averages results across varied prompts, which can obscure whether a test measures its intended skill or other factors.
- To validate BenchMIRT, researchers analyzed performance data from 100 LLMs across 16 benchmarks and more than 34K questions.
The dataset included six reasoning benchmarks like MMLU-Pro, GPQA, MATH, and BBH, along with ten safety benchmarks from the Olmo 3 safety suite such as HarmBench, StrongReject, WildJailbreak, and BBQ.
- The findings highlight critical nuances in how benchmarks evaluate models.
For example, the BBQ benchmark, which evaluates social bias and is typically categorized under safety, was found to align much more strongly with general reasoning.
- BenchMIRT applies multidimensional item response theory to measure prompt difficulty and discriminate model skills across evaluation tasks.
The framework evaluated performance data from 100 LLMs across 16 benchmarks and more than 34K questions.
- Analysis showed that social bias evaluations like BBQ aligned more closely with general reasoning capabilities than safety metrics.
Researchers at Ai2 created BenchMIRT to examine the specific capabilities driving performance scores in AI evaluations. Traditional benchmarking often averages results across varied prompts, which can obscure whether a test measures its intended skill or other factors. To address this, BenchMIRT adapts multidimensional item response theory from psychometrics, estimating prompt difficulty and model capabilities across prompt-level task data.
To validate BenchMIRT, researchers analyzed performance data from 100 LLMs across 16 benchmarks and more than 34K questions. The dataset included six reasoning benchmarks like MMLU-Pro, GPQA, MATH, and BBH, along with ten safety benchmarks from the Olmo 3 safety suite such as HarmBench, StrongReject, WildJailbreak, and BBQ. Without upfront instructions regarding benchmark categories, BenchMIRT consistently isolated safety and general reasoning as the two main operational dimensions.
The findings highlight critical nuances in how benchmarks evaluate models. For example, the BBQ benchmark, which evaluates social bias and is typically categorized under safety, was found to align much more strongly with general reasoning. Understanding these overlapping signals allows AI practitioners to design more precise evaluation frameworks and avoid misattributing model capabilities based solely on aggregate scores.
For more details please read the original article at Hugging Face.
Continue Learning
Comments
Comments appear only after moderation. Your email identifies your submission to the moderator and is never displayed here.
No approved comments yet.