EvoTrustRAG: Evolution-Aware Conflict Attribution and Evidence Handling for Reliable Retrieval-Augmented Generation

EvoTrustRAG is a training-free framework that, before RAG answer generation, attributes conflicting evidence to knowledge evolution, malicious manipulation, or unresolved uncertainty. It represents retrieved facts as a conflicting evidence graph, evaluates hypotheses using temporal relations, support structure, and auxiliary consistency, and projects them into a globally consistent explanation. In experiments, its average accuracy on benchmark-native conflicts reaches 81.4%, its attribution macro-F1 improves from 72.2% to 79.1% over the strongest baseline, and its error rate under the strongest coordinated attack drops from 31.2% to 16.0%.