Factuality metrics face trust issues in judging language models
Can We Trust the Judges? Validation of Factuality Evaluation Methods via Answer Perturbation
Computation and Language
Summary
Checking if language model answers are true is important but tricky. The paper finds that popular methods to judge truthfulness sometimes do not accurately notice small changes in answer correctness. The authors test different judging methods by deliberately corrupting correct answers to see how well the judges detect errors. They found that some step-by-step methods work better than those relying solely on other language models as judges. They also suggest a new, cheaper method that performs competitively.
What this means in practice
- •For nlp model developers: Improve the reliability of factuality checks in language model outputs by choosing better evaluation metrics informed by this paper’s meta-evaluation.
- •For fact-checking tool builders: Enhance automatic fact-checking tools by integrating pipeline-based factuality metrics that better track nuanced errors in generated answers.
Authors
Sarra Gharsallah, Adele Robaldo, Mariia Tokareva, Giovanni Gatti Pinheiro, Ilyana Guendouz, Raphaël Troncy, Paolo Papotti, Pietro Michiardi
Abstract
Evaluating the factual correctness of large language models (LLMs) is vital for many applications. But are our evaluation tools themselves trustworthy? Despite the rise of factuality-based metrics, their sensitivity and reliability remain underexplored. This paper introduces a meta-evaluation framework that systematically tests these metrics using controlled corruptions of gold standard answers. Our method generates ranked outputs with known degrees of degradation to probe how metrics capture nuanced changes in truthfulness. Our experiments reveal that pipeline-based methods, such as the RAGAS's factual correctness metric, better track degradation than LLM-as-judge approaches. We also propose a new variant of the factual correctness metric that provides a competitive and cost-efficient.