This paper explores the reliability of large language models (LLMs) when faced with different but equivalent question formulations. While LLMs often demonstrate strong accuracy on benchmark tests, their stability in applying this knowledge remains unclear.
We study four benchmarks and 13 models, finding that model outputs frequently depend on the exact wording of the prompt. Overall accuracy typically changes only modestly across paraphrases, but instance-level behavior is significantly less stable: for many questions, models alternate between correct and incorrect answers depending on phrasing, with mismatch rates exceeding 23%.
Focusing on questions that are answered correctly in their original form reveals even larger failures in terms of answer flip rates, indicating that single-prompt correctness is often a poor indicator of reliability. Nevertheless, we find that models often produce a correct answer for at least one paraphrase of a question, suggesting that the underlying knowledge is present but inconsistently retrieved.
Building on this observation, we show that a simple self-paraphrasing strategy can partially recover this latent knowledge and improve performance at inference time. Together, these findings suggest that standard accuracy metrics can mask substantial instability, and that evaluating consistency across equivalent inputs provides a clearer picture of LLM reliability.
Blogger's Review: This paper highlights the reliability issues of large language models when dealing with paraphrased inputs, stressing the limitations of accuracy metrics and proposing self-paraphrasing as a potential strategy for enhancing performance. This has significant implications for improving the practical utility of LLMs.