When Consistency Lies: Evaluating Prompt Paraphrasing for Black-Box Detection of Incorrect LLM Answers
Abstract
Can a model’s answers to repeated factual questions help identify when it is wrong? A wrong answer may be repeated consistently, so this study compared two black-box signals: answers to identical prompts and answers to meaning-preserving paraphrases. A fixed sample of 150 questions not used in the feasibility pilot from SimpleQA Verified was given to GPT-5 Nano and Claude Haiku 4.5. Each model answered each question in five independent API calls: three identical original prompts and two paraphrases (1,500 responses total). Original Run 1 (O1) was the target answer. Repeat consistency was the mean pairwise semantic agreement among O1, O2, and O3; paraphrase consistency used O1, P1, and P2. Error scores were defined as 1 minus consistency. Discrimination was measured using receiver operating characteristic area under the curve (ROC-AUC), with 95% confidence intervals from 10,000 question-level bootstrap resamples. After excluding 31 O1 nonanswers, 269 cases remained (37 correct and 232 incorrect). Repeat-consistency AUC was 0.751 (95% CI, 0.671-0.824), compared with 0.714 (0.618-0.801) for paraphrase consistency; the difference was -0.037 (95% CI, -0.117 to 0.043). Paraphrasing was not shown to improve overall discrimination. Of 232 incorrect answers, 52 (22.4%) remained perfectly consistent across identical repeats, and paraphrasing exposed disagreement in 14 of these 52 (26.9%). Paraphrases therefore revealed some errors that identical repeats missed, although superiority as an overall screening measure was not established.