Vis-Poison

This repository hosts the dataset accompanying Vis-Poison: Poisoning Visual Knowledge in Multimodal Retrieval-Augmented Generation.

The dataset contains 4,416 examples, each with a question, a correct answer, a target incorrect answer, and a pair of original and counterfactually edited images.

• Paper: arXiv:2608.20756
• GitHub repository: SWUFE-DB-Group/Vis-Poison
• Source dataset: WebQA

Data source and construction

Vis-Poison is derived from WebQA, introduced in WebQA: Multihop and Multimodal QA (CVPR 2022).

This release contains 4,416 selected and processed examples with paired original and counterfactually edited images. The Vis-Poison construction process adds target incorrect answers and editing annotations, including edit categories and instructions, for evaluating visual knowledge poisoning in multimodal retrieval-augmented generation.

This is a derived dataset, not the complete original WebQA benchmark. For the detailed selection and construction procedure, please refer to the Vis-Poison paper and code repository.

Citation

If you find our work useful or use it in your research, please consider citing our EMNLP 2026 paper:

@inproceedings{liang2026vispoison,
  title     = {Vis-Poison: Poisoning Visual Knowledge in Multimodal Retrieval-Augmented Generation},
  author    = {Liang, Rujin and Chen, Zhongpu and Lei, Yuhao and Miao, Xin},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}