When Does Uncertainty Measure RAG Hallucination Risk?

Accepted paper — no in-person presentation.

This paper remains accepted. No in-person poster presentation is scheduled.

Token uncertainty is often used to estimate hallucination risk in retrieval-augmented generation (RAG), but its meaning depends on the evidence supplied to the model. In RAG, perplexity should indicate hallucination risk when unsupported content is surprising under the retrieved context. It can fail when retrieval is sparse enough that the model answers from internal knowledge instead. We test this condition across four instruction-tuned language models and two question-answering datasets. After matching faithful and hallucinated answers within word-count bins, length-only classifiers fall to near chance, while sequence-level perplexity reaches 0.81--0.91 AUROC on controlled unsupported factual edits. Under sparse retrieval, the same uncertainty statistic can weaken or reverse; in reversal-prone cells, most recovery occurs after adding a second retrieved passage. Combining semantic drift with token-uncertainty features preserves controlled-benchmark discrimination, improves calibration, and remains above chance on human-annotated RAGTruth QA examples. These results show that uncertainty can measure RAG hallucination risk, but only under retrieval and evaluation conditions that make unsupported content unlikely under the supplied evidence.