REVEAL: Retrieval-Enhanced Verification for Multimodal Fact-Checking

Multimodal misinformation combines images and text to amplify false narratives, yet most fact-checking research addresses only textualclaims. The AVerImaTeC shared task introduces real-world image-text claims requiring sophisticated evidence retrieval. We present REVEAL (Retrieval-Enhanced Verification with Evidence Accumulation Loop), a system designed to overcome the “semantic gap,” defined as the disconnect between the neutral phrasing of claims and the adversarial vocabulary of debunking evidence. Unlike static baselines, REVEAL breaks down the verification task into an iterative context loop, integrating sparse and dense retrieval signals to aggressively target refuting evidence. We achieve a Verdict Accuracy of 23.6% and an Evidence Recall of 27.7% on the test set. Our results outperform the official baseline across all metrics, validating our hybrid retrieval strategy for complex multimodal verification.

Citation

A. Tariq, Y. Kementchedjhieva, "REVEAL: Retrieval-Enhanced Verification for Multimodal Fact-Checking," 2026, pp. 108-113.

Conference

Proceedings of the Ninth Fact Extraction and VERification Workshop (FEVER)

Source

Proceedings of the Ninth Fact Extraction and VERification Workshop (FEVER)

Publisher

Association for Computational Linguistics

DOI

10.18653/v1/2026.fever-1.8

Additional links

https://aclanthology.org/2026.fever-1.8/