MIA-Bench is a new benchmark designed to evaluate multimodal large language models (MLLMs) on their ability to strictly adhere to complex instructions. Our benchmark comprises a diverse set of 400 image-prompt pairs, each crafted to challenge the models’ compliance with layered instructions in generating accurate responses that satisfy specific requested patterns. We hope this benchmark not only serves as a tool for measuring MLLM adherence to instructions, but also guides future developments in MLLM training methods.
Figure 1: An example from MIA-Bench, featuring an image and a complex instruction to test models’ compliance with layered instructions that are compositional in nature. Responses from GPT-4v and InternVL-v1.5 are evaluated using GPT-4o as the judge.
Step 1:
- Install OpenAI API following its official document and prepare your API for GPT-4o.
Step 2:
- Run inference on the benchmark using your MLLM and store the responses in
jsonlformat. Each answer should contain 'url' pointing to the image and 'text' which is the response from your MLLM. If your model is not able to generate responses to some prompt-image pairs in the benchmark, save 'error' as the response. An example file can be found here. (Some urls may be unstable; it's more convenient to first download the images to a local folder before running inference. If you encounter problems with downloading images, please contact yqian22@apple.com)
Step 3:
- Load the inference result in
evaluation.ipynb. Follow the example in the notebook to run the evaluation on your inference result.
@misc{qian2024miabenchbetterinstructionfollowing,
title={MIA-Bench: Towards Better Instruction Following Evaluation of Multimodal LLMs},
author={Yusu Qian and Hanrong Ye and Jean-Philippe Fauconnier and Peter Grasch and Yinfei Yang and Zhe Gan},
year={2024},
eprint={2407.01509},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2407.01509},
}
Figure 2: An example with responses from four MLLMs and their evaluation scores.