Correcting Human Labels for Rater Effects in AI Evaluation: An Item Response Theory Approach

Researchers propose using psychometric modeling to correct systematic biases in human evaluations of AI systems, demonstrating how Item Response Theory can separate true AI output quality from rater behavior inconsistencies. The approach was tested on OpenAI's summarization dataset and showed improved reliability in measuring AI model performance.

• →Human evaluations of AI models contain systematic errors from rater effects like severity and centrality bias that distort conclusions.
• →Multi-faceted Rasch models can mathematically separate true AI output quality from individual rater behavioral patterns.
• →The approach was validated using OpenAI's summarization dataset, producing corrected estimates of summary quality.
• →Incorporating psychometric modeling into AI evaluation pipelines enables more principled use of human feedback data.
• →This methodology provides diagnostic insights into individual rater performance and reduces reliance on error-prone raw ratings.