ConRAG: Consensus-Driven Evidence Flow for Multi-Hop Question Answering
ConRAG is a consensus-driven framework designed to improve large language models' performance on complex multi-hop question answering by better selecting and transferring evidence. It combines multi-view retrieval signals—relation, entity, and text—into a unified ranking space using consensus fusion. This system also uses slot binding to pass intermediate information between reasoning steps, significantly boosting accuracy across multiple benchmarks.