DEFAME chooses tools to retrieve and evaluate text and image evidence. If the evidence is insufficient, it returns to planning to gather more. The final report contains a verdict and its supporting sources.
01 / Abstract
Abstract
DEFAME verifies image–text claims by selecting retrieval tools and reasoning over textual and visual evidence. Its modular, zero-shot pipeline uses a six-stage process to choose tools and search depth, evaluate the retrieved material, and produce a structured fact-checking report. The evaluation covers VERITE, AVeriTeC, MOCHEG, and the new ClaimReview2024+ benchmark, whose claims postdate the backbone model’s knowledge cutoff. On ClaimReview2024+, DEFAME outperforms the reported GPT-4o baselines.
02 / Contributions
Contributions
01
End-to-end multimodal verification
Handles images in both claims and retrieved evidence while producing a structured, evidence-grounded report.
02
Dynamic investigation
Lets the model choose tools and search depth instead of applying one fixed retrieval recipe to every claim.
03
Evaluation beyond the knowledge cutoff
Introduces ClaimReview2024+, with claims published after the backbone model's knowledge cutoff.
03 / Citation
Citation
Braun, T., Rothermel, M., Rohrbach, M., & Rohrbach, A. (2025). DEFAME: Dynamic Evidence-based FAct-checking with Multimodal Experts. Proceedings of the 42nd International Conference on Machine Learning, 267, 5383–5417.
BibTeX+
@inproceedings{braun2025defame,
title = {{DEFAME}: Dynamic Evidence-based {FA}ct-checking with Multimodal Experts},
author = {Tobias Braun and Mark Rothermel and Marcus Rohrbach and Anna Rohrbach},
booktitle = {Proceedings of the 42nd International Conference on Machine Learning},
volume = {267},
pages = {5383--5417},
year = {2025},
url = {https://proceedings.mlr.press/v267/braun25b.html}
}