X-MADAM-RAG: Diagnosing and Handling Chinese-English Evidence Conflict in Retrieval-Augmented Generation

The article introduces X-MADAM-RAG, a novel pipeline designed to address evidence conflict in retrieval-augmented generation (RAG) systems, particularly in multilingual contexts involving Chinese and English. It utilizes the X-RAMDocs-ZHEN benchmark, comprising 300 examples to diagnose these conflicts, and demonstrates that X-MADAM-RAG achieves a strict accuracy of 0.9667 and a conflict-aware success rate of 0.9767 when tested with the Qwen2.5-7B-Instruct model. This work highlights critical limitations in document-level extraction processes and positions both X-RAMDocs-ZHEN and X-MADAM-RAG as tools for diagnosing evidence conflicts rather than as solutions for general hallucination detection.