Elamur Research⌕
☰
home / papers / 2026 / AIC CTU@AVerImaTeC: dual-retriever RAG for image-text fact checking

AIC CTU@AVerImaTeC: dual-retriever RAG for image-text fact checking

Paper · 2026 · Herbert Ullrich and Jan Drchal

Summary

Verifying claims that pair a statement with pictures means judging the text and also whether the images are authentic and shown in the right context. For the AVerImaTeC shared task, the CTU AI Center authors adapt their earlier AVeriTeC RAG pipeline by adding a second retriever.

  • Text branch: the offline knowledge store is split into 2048-character chunks and embedded with mxbai-embed-large-v1. It is searched with exact k-NN in FAISS, and the hits are diversified with MMR.
  • Image branch: each claim image goes to Google Lens through the Serper API. The returned pages are scraped with Firecrawl, and pages published after the claim or blocked from scraping are removed.
  • Generation: both source sets, plus BM25-chosen few-shot examples from the train set, go into a single multimodal GPT-5.1 call. It outputs evidence, a verdict and a justification.

The system finished 3rd and clearly beat the iterative agentic baseline, with generation costing about $0.013 per claim. Because it is cheap, modular and easy to rebuild, the authors offer it as an accessible baseline and release code, prompts and vector stores.

Key points

  • Test-split scores: question 0.81 (near the best), evidence 0.35, combined verdict 0.35, justification 0.3. Metrics use Ev2R recall with an LLM judge.
  • Retrieval settings: k = 20 neighbours, reduced by MMR (λ = 0.8) to 7 text sources. Up to 30 Google Lens hits per image are capped at 9 usable sources.
  • Costs: Serper search is about $0.003 per image and Firecrawl is $0.006 per page; the free Trafilatura library is suggested as a substitute. GPT-5.1 via the Batch API used about 11K input and 1150 output tokens per claim.
  • The authors identify the evidence score as the main weakness, partly because the system writes evidence as QA pairs, which does not match AVerImaTeC's format. Generating declarative evidence in an ablation raised the evidence score by about 2 points.
  • Stated limitations:
    • Google Lens often returns nothing for older or explicit images.
    • The pipeline depends on proprietary services.
    • There is no text-only vs. text+RIS ablation.
    • The data are heavily skewed toward Refuted (95% of train claims).

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 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 worksSocial Sciences · 1,772 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 Herbert Ullrich and Jan Drchal. 2026. AIC CTU@AVerImaTeC: dual-retriever RAG for image-text 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-1.11

@inproceedings{ullrich2026aic,
  title = {{AIC CTU@AVerImaTeC: dual-retriever RAG for image-text fact checking}},
  author = {Herbert Ullrich and Jan Drchal},
  booktitle = {Proceedings of the Ninth Fact Extraction and VERification Workshop (FEVER)},
  year = {2026},
  doi = {10.18653/v1/2026.fever-1.11},
  url = {https://doi.org/10.18653/v1/2026.fever-1.11}
}

References

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

  1. Mubashara Akhtar, Rami Aly, Yulong Chen, Zhenyun Deng, Michael Schlichtkrull, Chenxi Whitehouse, and Andreas Vlachos. 2025. The 2nd automated verification of textual claims (AVeriTeC) shared task: Open-weights, reproducible and efficient systems. In Proceedings of the Eighth Fact Extraction and VERi… find
  2. Mubashara Akhtar, Michael Schlichtkrull, and Andreas Vlachos. 2024. Ev2r: Evaluating evidence retrieval in automated fact-checking. find
  3. Stephen Robertson, S. Walker, S. Jones, M. M. Hancock-Beaulieu, and M. Gatford. 1995. Okapi at trec-3. In Overview of the Third Text REtrieval Conference (TREC-3).
  4. Adrien Barbaresi. 2020. htmldate: A Python package to extract publication dates from web pages. Journal of Open Source Software, 5(51):2439. find
  5. 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. find
  6. Jaime Carbonell and Jade Goldstein. 1998. The use of mmr, diversity-based reranking for reordering documents and producing summaries. In Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’98, page 335–336, New York, NY, USA.… find
  7. Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The faiss library.
  8. Max Glockner, Yufang Hou, and Iryna Gurevych. 2022. Missing counter-evidence renders NLP fact-checking unrealistic for misinformation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5916–5936, Abu Dhabi, United Arab Emirates. Association for Computat… find
  9. Michael Schlichtkrull, Yulong Chen, Chenxi Whitehouse, Zhenyun Deng, Mubashara Akhtar, Rami Aly, Zhijiang Guo, Christos Christodoulopoulos, Oana Cocarascu, Arpit Mittal, James Thorne, and Andreas Vlachos. 2024. The automated verification of textual claims (AVeriTeC) shared task. In Proceedings of th… find
  10. Herbert Ullrich and Jan Drchal. 2025. AIC CTU@FEVER 8: On-premise fact checking through long context RAG. In Proceedings of the Eighth Fact Extraction and VERification Workshop (FEVER), pages 274–280, Vienna, Austria. Association for Computational Linguistics. find
  11. Herbert Ullrich, Tomáš Mlynář, and Jan Drchal. 2024. AIC CTU system at AVeriTeC: Re-framing automated fact-checking as a simple RAG task. In Proceedings of the Seventh Fact Extraction and VERification Workshop (FEVER), pages 137–150, Miami, Florida, USA. Association for Computational Linguistics. find
  12. Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2022. A survey on automated fact-checking. Transactions of the Association for Computational Linguistics, 10:178–206.
  13. Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535–547.
  14. Sean Lee, Aamir Shakir, Darius Koenig, and Julius Lipp. 2024. Open source strikes bread - new fluffy embeddings model. find
  15. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Proceedings of the 34th Internation…
  16. Xianming Li and Jing Li. 2024. AoE: Angle-optimized embeddings for semantic textual similarity. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1825–1839, Bangkok, Thailand. Association for Computational Linguistics. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ullrich-2026-aic-ctu-averimatec-dual-retriever/