When Does Uncertainty Measure RAG Hallucination Risk?

A new arXiv paper (2610.22615v1) finds that uncertainty measures hallucination risk in retrieval-augmented generation only when retrieved evidence constrains the answer. Testing four instruction-tuned models on HotpotQA and SQuAD, the authors report that after matching faithful and hallucinated answers within word-count bins, length-only classifiers fall near chance while perplexity reaches 0.81–0.91 AUROC on targeted unsupported edits, and that detection can weaken with only the top-ranked passage before recovering after a second passage is added. A combined uncertainty and semantic-drift score improves calibration, and the measurements remain above chance on RAGTruth QA.

arXiv:2610.22615v1 Announce Type: new Abstract: Uncertainty can measure hallucination risk in retrieval-augmented generation RAG , but only when retrieved evidence constrains the answer. When evidence is present, unsupported factual claims become less likely; when retrieval coverage is poor, the model can answer from internal knowledge and assign high probability to unsupported claims. We test this across four instruction-tuned models on HotpotQA and SQuAD. After matching faithful and hallucinated answers within word-count bins, length-only classifiers fall near chance, while perplexity reaches 0.81--0.91 AUROC on targeted unsupported edits. Varying retrieval depth shows why the score changes meaning: detection can weaken with only the top-ranked passage, then recover after adding a second passage. A combined uncertainty and semantic-drift score improves calibration, and measurements remain above chance on RAGTruth QA. Results support a conditional view of uncertainty as a useful measure of RAG hallucination risk.