Luca Gioffré

PhD Student in NLP at Sapienza, University of Rome :it:

1 minute read

📄 Read the full paper on ACL Proceedings ACL or on arXiv arXiv

Can we truly trust LLM evaluation scores? 🤔

Multiple-Choice Question Answering is one of the most widely used benchmarks for assessing LLMs across domains like math, science, commonsense reasoning and general knowledge. Since answer choices are predefined, evaluating models seems straightforward. However, our latest research reveals that current evaluation strategies introduce inconsistencies that may lead to incorrect conclusions about model performance.

In our paper, “Right Answer, Wrong Score: Uncovering the Inconsistencies of LLM Evaluation in Multiple-Choice Question Answering”, we conduct a systematic analysis of how different answer extraction methods influence reported accuracy. Our study highlights several critical challenges that affect the reliability of LLM evaluation.

Our Key Findings:

  1. Traditional evaluation strategies underestimate LLM performance
  2. A trade-off exists between enforcing format constraints (for easy evaluation) and allowing free-form text (for better reasoning)
  3. LLM-based answer extractors are prone to systematic errors

We show that the way in which answers are extracted can significantly impact reported accuracy, raising concerns about the robustness of existing MCQA benchmarks. Our findings highlight the need for more reliable and standardized evaluation strategies to ensure fair model comparison.

👏 Huge thanks to all my co-authors!

~ Francesco Maria Molfese, Luca Moroni, Luca Gioffré, Alessandro Scirè, Simone Conia and Roberto Navigli.