Authors: H Ullrich, J Drchal
Venue: arXiv preprint arXiv:2602.15190
Year: 2026
Citations: N/A
Links
- arXiv: 2602.15190
Abstract
This paper introduces a dual-retriever RAG system for multimodal fact-checking, combining textual and image-aware evidence retrieval before verdict generation. The architecture is built to reduce hallucinations by forcing stronger evidence alignment across modalities.
Resources
- Video: TODO
- Slides: TODO
- Code: TODO
- Dataset: TODO
Notes
A highlight of this project was a bronze-medal finish in AVeriTeC 2. Beyond leaderboard performance, we are proud that the system remained interpretable: retrieval traces made it easier to analyze failures and iterate quickly.