Logic Matters in Lightweight Hallucination Classification for RAG System

We propose a lightweight, modular framework for hallucination detection in Retrieval-Augmented Generation (RAG) systems, addressing the critical challenge where logical dependencies span across fragmented retrieval results. Through graph-based semantic evidence aggregation, which captures the implicit logical structure by clustering semantically coherent segments across retrieved documents via betweenness centrality, our approach enables small NLI models to handle multihop reasoning without task-specific training. We present two deployment configurations: a resource-efficient variant (≈0.5B parameters) achieving 82.4% accuracy on HotPotQA-Derived at 85 ms latency, outperforming all sub-1B baselines by over 30%; and a higheraccuracy variant (≈1.5B parameters) reaching 85.6%, surpassing 11B TrueTeacher while being 7× smaller and 1.7× faster. Experiments with six NLI discriminator models show consistent gains of +6.7%-+29.9%, confirming that graph-based evidence aggregation is NLIagnostic and the primary performance driver. We also contribute HotPotQA-Derived, a new multi-hop hallucination benchmark preserving separate retrieved documents for systematic evaluation.