Beyond epistemic collapse: disagreement-aware scientific RAG

Our paper has been accepted to EMNLP 2026, a CORE A* conference in NLP: Beyond Epistemic Collapse: Disagreement-Aware Scientific Retrieval-Augmented Generation.

What we built

We introduce EVIRAG, a disagreement-aware RAG framework designed to prevent LLMs from collapsing conflicting scientific evidence into a single answer. We also introduce EVIRAG-BENCH, a 1,250-query benchmark spanning five scientific domains.

Results

Our approach substantially improves contradiction recall and viewpoint coverage over standard RAG.

The benchmark and pipeline are open: EVIRAG-Bench on GitHub.

Thanks

Big thanks to Krishang Sharma for the work and ideas throughout this project, and to Sonia Khetarpaul for all the guidance and support. Looking forward to presenting this in Budapest.