Same Question, Different Answers: Evaluating LLM Reliability Beyond Accuracy

Summary

The authors demonstrate that large language models (LLMs) frequently flip between correct and incorrect answers when questions are merely rephrased, across four benchmarks and 13 models. While overall accuracy rates change only modestly, instance-level inconsistency rates exceed 23%.

Why it matters

If a model's response depends on superficial phrasing rather than stable knowledge, one-shot accuracy benchmarks overestimate real-world reliability. This is critical in any deployment scenario where users naturally rephrase their queries. This finding necessitates the use of consistency metrics alongside accuracy in evaluation pipelines. The observation that knowledge is often present but retrieved inconsistently points to practical solutions applicable at inference time, rather than purely architectural changes.

Highlights

• Although overall accuracy rates remain nearly unchanged, instance-level answer flip rates exceed 23% under meaning-preserving rephrasings.
• Accuracy in the original prompt is a weak indicator of reliability. When focusing on questions that were initially answered correctly, answer flip rates are found to be even higher.
• A simple self-rephrasing strategy applied at inference time partially recovers latent knowledge and improves performance.

Critical take

The 23% figure is an upper bound ('exceeding 23%'), not an average. The study is limited to domains with binary accuracy, such as factual question-answering and mathematical reasoning. Consequently, the instability observed here appears sharper than what would likely be seen in open-ended generation. Additionally, ensuring that rephrasings truly preserve meaning is inherently difficult. It is explicitly stated that the self-rephrasing correction is only partial, meaning it falls short of resolving core-level retrieval inconsistencies.