CoCR-RAG: Concept-Oriented Context Reconstruction Outperforms Existing RAG Methods on Web Q&A Benchmarks
arXiv·medium signal
CoCR-RAG tackles multi-source information fusion in RAG through linguistically grounded concept-level integration rather than passage-level concatenation. On PopQA and EntityQuestions benchmarks, CoCR-RAG significantly outperforms existing context-reconstruction methods by decomposing retrieved passages into atomic concepts and reconstructing coherent context around query-relevant concept clusters.