Reproduce “BayesRAG: Probabilistic Mutual Evidence Corroboration for Multimodal Retrieval-Augmented Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xuan Li, Yining Wang, Haocai Luo, Shengping Liu, Jerry Liang, Ying Fu, Weihuang, Jun Yu and 1 more
BayesRAG answers questions about long documents by ranking tuples of text, visual content, and page-structure evidence. It combines modality scores with Dempster–Shafer evidence theory and uses a consistency prior to estimate whether retrieved pieces belong together. On MMLongBench-Doc, with GPT-4o-mini, it scores 44.1%, versus 41.5% for RAG-Anything; this is an end-to-end answer metric, not proof of causation.
Key insights
- Explicitly modeling semantic consistency and resolving cross-modal conflicts through a theoretically grounded probabilistic framework significantly enhances the trustworthiness and robustness of multimodal RAG.
- Integrating both semantic (Knowledge Graph) and spatial (document layout) priors enables a comprehensive understanding of evidence coherence, effectively recovering structural information lost by traditional chunking.
- Probabilistic evidence fusion acts as an efficient soft reranker for visual candidates, mitigating 'high-similarity hallucinations' and boosting effective recall by identifying mutually corroborating evidence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.