Running summarizing tools on a new article

Abstract: Retrieval-Augmented Generation (RAG) has significantly improved the factualreliability of Large Language Models (LLMs) by grounding generated responses inexternally retrieved evidence. Despite these advances, hallucinations remain acritical challenge due to retrieval failures, long-context limitations, andparametric memory override, where language models generate unsupportedinformation despite the presence of relevant evidence. Existing post-hochallucination detection methods commonly evaluate entire responses using staticconfidence aggregation strategies, making them susceptible to error dilution andpoor confidence calibration. This paper proposes Adaptive Hierarchical Confidence Estimation (AHCE), alightweight, model-agnostic framework for claim-level hallucination detection inRetrieval-Augmented Generation systems. Unlike conventional response-levelverification methods, AHCE decomposes generated responses into atomic factualclaims and performs fine-grained evidence verification for each claimindependently. To support long-context retrieval scenarios, the frameworkintroduces span-level context windowing, enabling efficient Natural LanguageInference (NLI) verification without requiring expensive long-context languagemodels. AHCE integrates multiple complementary confidence signals, including semanticentailment, lexical grounding, embedding similarity, and native generatoruncertainty derived from token-level log probabilities. Rather than relying onmanually selected confidence weights, these heterogeneous signals are fused by asupervised meta-classifier that learns adaptive confidence estimation acrossdiverse query types. Finally, probability calibration produces reliableconfidence estimates suitable for deployment-oriented abstention decisions. AHCE unifies claim-level verification, evidence localization, adaptive fusion,and calibrated uncertainty estimation in a single post-hoc pipeline that requiresno fine-tuning of the underlying model. We empirically validate AHCE on theRAGTruth hallucination-detection benchmark under a strict fit-on-development /evaluate-on-held-out protocol, and further show that the efficiency of thehierarchical Dual-NLI cascade generalizes to a second, independently constructedbenchmark (HaluEval): across both datasets the confidence-triggered cascadematches the accuracy of always-run single- and dual-NLI ensembles withinstatistical noise while using roughly 3x less NLI compute. In the interest ofscientific transparency, we also report the components whose gains do notgeneralize out-of-sample (Section 4).