VisRecall

This repository contains the VisRecall benchmark, introduced in Traveling Across Languages: Benchmarking Cross-Lingual Consistency in Multimodal LLMs.

Dataset Description

Imagine a tourist finished their journey in Japan and came back to France, eager to share the places they visited with their friends. When portraying these experiences, the visual information they convey is inherently independent of language, meaning that descriptions created in different languages should ideally be highly similar. This concept extends to MLLMs as well. While a model may demonstrate decent consistency in VQA tasks, any inconsistency in generation tasks would lead to a biased user experience (i.e., a knowing vs saying distinction). To assess the cross-lingual consistency of "visual memory" in MLLMs, we introduce VisRecall, a multilingual benchmark designed to evaluate visual description generation across 450 landmarks in 9 languages.

The dataset contains the following fields:

Field Name	Description
landmark_id	Unique identifier for the landmark in the dataset.
domestic_language_code	ISO 639 language code of the official language spoken in the country where the landmark is located.
language_code	ISO 639 language code of the prompt.
country_code	ISO country code representing the location of the landmark.
landmark_name	Name of the landmark used for evaluation.
prompt_idx	Index of the prompt used. Each language includes two distinct prompts.

Additionally, the following files are necessary for running evalutaion:

File Name	Description
images.tar.gz	Compressed archive containing images of landmarks, used for CLIPScore calculation.
images_list.json	List of image file paths included in the dataset.
landmark_list.json	Metadata for each landmark, including IDs, names, etc.

Evaluation

Please refer to our GitHub repository for detailed information on the evaluation setup.

Citation

@misc{wang2025travelinglanguagesbenchmarkingcrosslingual,
      title={Traveling Across Languages: Benchmarking Cross-Lingual Consistency in Multimodal LLMs}, 
      author={Hao Wang and Pinzhi Huang and Jihan Yang and Saining Xie and Daisuke Kawahara},
      year={2025},
      eprint={2505.15075},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.15075}, 
}