Tag
This paper introduces X-RAMDocs-ZHEN, a controlled Chinese-English benchmark for diagnosing evidence conflicts in retrieval-augmented generation, and X-MADAM-RAG, an interpretable pipeline. While the pipeline outperforms baselines on the original benchmark, it shows limitations under a naturalized stress test, highlighting document-level extraction as a key bottleneck.
This paper evaluates six open-weight LLMs on biomedical QA under conflicting evidence conditions, revealing accuracy drops and prediction flips, and proposes a conflict-aware abstention score that improves selective accuracy.