Elamur Research⌕
☰
home / papers / 2026 / REVEAL: Retrieval-Enhanced Verification for Multimodal Fact-Checking

REVEAL: Retrieval-Enhanced Verification for Multimodal Fact-Checking

Paper · 2026 · Amina Tariq and Yova Kementchedjhieva

Summary

Most automated fact-checking work targets text-only claims, but much real misinformation pairs images with text. This FEVER 2026 paper presents REVEAL, a system for the AVerImaTeC shared task, in which real-world image-text claims must be verified against a fixed knowledge store. The authors argue that the official baseline (BM25 plus CLIP) suffers from a "semantic gap": claims tend to be worded neutrally, while refuting articles use charged terms such as "hoax" or "false." REVEAL keeps the baseline's question-planning agent but replaces its retrieval tools. It adds HyDE-generated pseudo-evidence and a weighted multi-query scheme, fused with a modified Reciprocal Rank Fusion that favours debunking-oriented queries. It also applies dense reranking with SFR-Embedding-2_R, boosts trusted fact-checking domains, and uses SigLIP2 for image search. An accumulation loop summarizes earlier evidence for the Qwen2.5-VL verifier. On the test set, REVEAL roughly doubles the baseline's verdict accuracy. The results suggest that evidence quality, rather than the reasoning ability of the multimodal model, is the main bottleneck in this task.

Key points

  • Test results: Evaluated on the 352 claims in AVerImaTeC's test split, REVEAL achieves 23.6% Verdict Accuracy and an Evidence Score of 0.277. The official baseline scores 11.4% and 0.171, and REVEAL places 6th on the leaderboard.
  • Query weighting: Four parallel query types are fused with RRF (k = 60), each with its own weight:
    • debunking-term queries: 3.0
    • LLM question: 2.0
    • raw claim: 1.5
    • HyDE text: 1.0
  • Reranking and domain boost: SFR-Embedding-2_R reranks the top 500 BM25 candidates and keeps 15. Sources such as Snopes, PolitiFact and Reuters receive a 1.5× score boost.
  • Ablation (dev set, 152 claims): HyDE gives the largest evidence gain (0.245 → 0.279), while swapping CLIP for SigLIP2 alone helps only marginally. The full system reaches 30.3% verdict accuracy.
  • Limitations:
    • The backbone is Qwen2.5-VL-7B-Instruct, and complex claims take about 3 minutes each, mainly because of reranking cost.
    • Observed errors include over-cautious "not enough evidence" verdicts, overly aggressive refutations, and confusion between visually similar locations.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputer Vision and Pattern Recognition · 2,869 worksSocial Sciences · 1,772 worksInformation Systems · 1,437 worksSociology and Political Science · 577 works

🧭 Topics 2Research subjects and languages the work is about.Vision & language · 657 worksInformation retrieval & search · 473 works

Cite

Cite as Amina Tariq and Yova Kementchedjhieva. 2026. REVEAL: Retrieval-Enhanced Verification for Multimodal Fact-Checking. Proceedings of the Ninth Fact Extraction and VERification Workshop (FEVER).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.fever-1.8

@inproceedings{tariq2026reveal,
  title = {{REVEAL: Retrieval-Enhanced Verification for Multimodal Fact-Checking}},
  author = {Amina Tariq and Yova Kementchedjhieva},
  booktitle = {Proceedings of the Ninth Fact Extraction and VERification Workshop (FEVER)},
  year = {2026},
  doi = {10.18653/v1/2026.fever-1.8},
  url = {https://doi.org/10.18653/v1/2026.fever-1.8}
}

References

5 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 others. 2025. Qwen2.5-vl technical report. Preprint, arXiv:2502.13…
  2. Rui Cao, Zifeng Ding, Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2025. Averimatec: A dataset for automatic verification of image-text claims with evidence from the web. arXiv preprint arXiv:2505.17978. supervision. In Proceedings of the 38th International Conference on Machine Learnin… — [arXiv:2505.17978](https://arxiv.org/abs/2505.17978) find
  3. Stephen Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval, 3:333–389. find
  4. Shafiq Rayhan Joty Caiming Xiong Yingbo Zhou Semih Yavuz Rui Meng*, Ye Liu*. 2024. Sfr-embedding-2: Advanced text embedding with multi-stage training. find
  5. Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier Hénaff, Jeremiah Harmsen, Andreas Steiner, and Xiaohua Zhai. 2025. Siglip 2: Multilingual vision-language encoders with improv… find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tariq-2026-reveal-retrieval-enhanced-verification-multimodal/