When Does Uncertainty Measure RAG Hallucination Risk?

Rajeel Ahmad Ansari; Tautik Agrahari

cs.CLOct 15, 2026

When Does Uncertainty Measure RAG Hallucination Risk?

Authors: Rajeel Ahmad Ansari, Tautik Agrahari

Organizations: Independent Researcher, Delhi, India · Independent Researcher, Bengaluru, India

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.

Figures & tables

Model	Dataset	Len.	PPL	Drift+ [-1pt]Uncertainty	Best
		AUROC			base.
Phi-3-mini	HotpotQA	0.502	0.811	0.842	0.741
Phi-3-mini	SQuAD	0.504	0.881	0.859	0.695
Gemma-3-4B	HotpotQA	0.509	0.908	0.911	0.839
Gemma-3-4B	SQuAD	0.519	0.871	0.868	0.753
Llama-3.1-8B	HotpotQA	0.517	0.883	0.846	0.740

Table 1: Length control removes the answer-length shortcut, while perplexity and Drift+Uncertainty remain discriminative. PPL denotes sequence-level perplexity; Best baseline is the strongest of BERTScore, NLI, and LettuceDetect.

Figure 1: Retrieval-depth stress test. In reversal-prone cells, perplexity recovers sharply from one to two retrieved passages, showing that likelihood-based scores depend on retrieval coverage.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Dataset	$n$	Non-H	Len. AUROC
Phi-3-mini	HotpotQA	2208	25.75	0.502
Phi-3-mini	SQuAD	1882	25.50	0.504
Gemma-3-4B	HotpotQA	1414	26.79	0.509
Gemma-3-4B	SQuAD	1238	27.78	0.519
Llama-3.1-8B	HotpotQA	774	30.66	0.517
Llama-3.1-8B	SQuAD	746	30.04	0.530

Appendix

Table 2: Length control after bin matching. The length-only classifier is near chance in every controlled cell. Non-H denotes mean non-hallucinated answer length; matched hallucinated lengths are within the same bins.

Scoring model	PPL	Mean P	Drift	Drift+ [-1pt]Uncertainty
Phi-3-mini	0.606	0.631	0.519	0.648
Gemma-3-4B	0.515	0.636	0.519	0.650
Llama-3.1-8B	0.599	0.613	0.519	0.629
Qwen2.5-14B	0.615	0.646	0.519	0.644
Mean	0.584	0.632	0.519	0.643

Appendix

Table 3: External validation on 875 RAGTruth QA examples. Mean P denotes mean token probability; Drift denotes semantic drift. The combined score remains above chance for every scoring model.

Metric	PPL	Drift+ [-1pt]Uncertainty
Mean ECE	0.072	0.035
Mean Brier score	0.187	0.150
Median recall at $\geq$ 95% precision	0.366	0.297

Appendix

Table 4: The combined score improves probability calibration, while raw perplexity has slightly higher recall at the high-precision threshold.

Figure 2: Higher unsupported-answer rates correspond to lower

k=1

perplexity AUROC across random samples (Spearman

\rho=-0.805

). Colors and marker shapes distinguish scoring models and datasets.

Cell	$k$	$n$	Unsup.	PPL AUROC (95% CI)
Natural stress
Gemma-3-4B + HotpotQA	1	100	71	0.285 [0.180, 0.400]
Qwen2.5-14B + SQuAD	1	100	93	0.200 [0.069, 0.398]
Seeded vary- $k$
Qwen2.5-14B + SQuAD	2	100	43	0.757 [0.658, 0.847]
Qwen2.5-14B + SQuAD	3	100	43	0.759 [0.662, 0.849]

Appendix

Table 5: Retrieval-stress sample sizes and bootstrap intervals. Natural stress rows report the

k=1

cells highlighted in the main text; seeded rows correspond to Figure 1 .

Cell	$k$	Agreement	$\kappa$
Natural stress cells	1	0.998	0.995
Qwen2.5-14B + SQuAD	2–4	0.963	0.926
Gemma-3-4B + HotpotQA	2–4	0.913	0.798
Llama-3.1-8B + HotpotQA	2–4	0.940	0.871
Overall vary- $k$	2–4	0.939	0.865

Appendix

Table 6: Cross-judge agreement for retrieval-shift labels. The natural stress row pools the five

k=1

cells used for the natural retrieval-shift analysis.

Hallucination in RAG

See all 19 papers

Facet-Level Tracing of Evidence Uncertainty and Hallucination in RAG
Apr 10, 2026Passant Elchafei, Monorama Swain, Shahed Masoudian +1Hallucination in RAG RAG Evaluation
Johannes Kepler University Linz, Institute of Computational Perception, Linz, Austria · Linz Institute of Technology, Artificial Intelligence Lab, Austria
Evidence Graph Consistency in Retrieval-Augmented Generation: A Model-Dependent Analysis of Hallucination Detection
Jun 4, 2026Jianru ShenRetrieval-Augmented Generation Hallucination in Language Models
University of Montana Missoula, MT, USA
SURE-RAG: Sufficiency and Uncertainty-Aware Evidence Verification for Selective Retrieval-Augmented Generation
May 5, 2026Jingxi Qiu, Zeyu Han, Cheng HuangRetrieval-Augmented Generation LLM Answer Verification
ZenWeave AI · Georgetown University · Southern Methodist University

Apr 10, 2026J/K move · Enter open · S save

Facet-Level Tracing of Evidence Uncertainty and Hallucination in RAG

Passant Elchafei, Monorama Swain, Shahed Masoudian, Markus Schedl

Johannes Kepler University Linz, Institute of Computational Perception, Linz, Austria · Linz Institute of Technology, Artificial Intelligence Lab, Austria

Hallucination in RAG RAG Evaluation Evidence-Grounded Generation

Open paper AskPDF arXiv

Retrieval-Augmented Generation (RAG) aims to reduce hallucination by grounding answers in retrieved evidence, yet hallucinated answers remain common even when relevant documents are available. Existing evaluations focus on answer-level or passage-level accuracy, offering limited insight into how evidence is used during generation. In this work, we introduce a facet-level diagnostics framework for QA that decomposes each input question into atomic reasoning facets. For each facet, we assess evidence sufficiency and grounding using a structured Facet x Chunk matrix that combines retrieval relevance with natural language inference-based faithfulness scores. To diagnose evidence usage, we analyze three controlled inference modes: Strict RAG, which enforces exclusive reliance on retrieved evidence; Soft RAG, which allows integration of retrieved evidence and parametric knowledge; and LLM-only generation without retrieval. Comparing these modes enables thorough analysis of retrieval-generation misalignment, defined as cases where relevant evidence is retrieved but not correctly integrated during generation. Across medical QA and HotpotQA, we evaluate three open-source and closed-source LLMs (GPT, Gemini, and LLaMA), providing interpretable diagnostics that reveal recurring facet-level failure modes, including evidence absence, evidence misalignment, and prior-driven overrides. Our results demonstrate that hallucinations in RAG systems are driven less by retrieval accuracy and more by how retrieved evidence is integrated during generation, with facet-level analysis exposing systematic evidence override and misalignment patterns that remain hidden under answer-level evaluation.

Abstract

Hallucination in RAG

Explore similar work

Facet-Level Tracing of Evidence Uncertainty and Hallucination in RAG

Evidence Graph Consistency in Retrieval-Augmented Generation: A Model-Dependent Analysis of Hallucination Detection

SURE-RAG: Sufficiency and Uncertainty-Aware Evidence Verification for Selective Retrieval-Augmented Generation