Grounding Probes: Generator-Independent Hallucination Detection from Observer Model Hidden States
A separate observer model can flag ungrounded RAG answers without access to the generator’s internals.
The paper’s “Grounding Probe” uses logistic regression on mean-pooled middle-layer hidden states from an observer LM reading the context, question, and response in one pass. It reports 0.879–0.894 AUROC on RAGTruth across four observers after training on 15,090 annotated responses. Averaging it with a supervised span detector reaches 0.924 AUROC and 0.820 [email protected], 0.060 AUROC above the span detector alone. The authors say one probe transfers across six generators, with removing a generator costing about 0.02 AUROC in hold-out controls. ArXiv · AI/CL/LG's note
The paper’s “Grounding Probe” uses logistic regression on mean-pooled middle-layer hidden states from an observer LM reading the context, question, and response in one pass. It reports 0.879–0.894 AUROC on RAGTruth across four observers after training on 15,090 annotated responses. Averaging it with a supervised span detector reaches 0.924 AUROC and 0.820 [email protected], 0.060 AUROC above the span detector alone. The authors say one probe transfers across six generators, with removing a generator costing about 0.02 AUROC in hold-out controls. ArXiv · AI/CL/LG's note
score 5