Work

Do LLMs Know When They Are Wrong? Resampling vs. Paraphrase Disagreement as Black-Box Error Signals

Year: 2026
Type: preprint
Abstract: Disagreement across paraphrases flagged wrong answers better than disagreement across resamples at the same call budget, but both signals are weak because most errors are confident. On DeepSeek-chat with 1,600 MMLU and 1,600 TriviaQA questions, AUROC for detecting errors with 3 extra answers was 0.695 (paraphrase) vs 0.646 (resample) on MMLU (difference +0.049, 95% CI 0.011 to 0.089) and 0.785 vs 0.668 on TriviaQA (+0.117, CI 0.065 to 0.169). Both differences remain significant after Holm correction. About 70% of errors are confident (at least 4 of 5 samples agree on the same wrong answer). Resampling flags 10% of them and paraphrasing 40%, at a higher false-alarm rate (9.4% vs 4.0%). Caveats: one model, DeepSeek writes and checks its own paraphrases, the MMLU gap shrinks to +0.031 (CI -0.009 to 0.071) when questions with missing paraphrases are kept, and 27% to 30% of audited "errors" were label noise or ambiguous questions. These are preliminary results. Preliminary results. Single model (DeepSeek-chat), 1,600 MMLU and 1,600 TriviaQA questions. Hand audits were first drafted with an LLM (Claude) and entered by the author; the report text was written with LLM assistance.
Language: English
Open Access status: green
OpenAlex ID: W7233739253
Cites:
Cited by:
Related to: 10