Trust Before Fusion: QIMG-7 and Source-Aware Resolution for Polluted Multimodal RAG

The paper introduces QIMG-7, a benchmark testing Retrieval-Augmented Generation (RAG) robustness against multimodal pollution, where text and images are intentionally corrupted. It proposes Source-Aware Trust Resolution (SATR), a training-free method that selects or falls back on candidate answers based on the reliability of their sources. Results show that source-aware trust resolution outperforms unconditional fusion, proving selective trust is better than blindly combining all evidence.

Open the full episode

More from AI Radio