Home

Awesome

Holistic Analysis of Hallucination in GPT-4V(ision): Bias and Interference Challenges

Chenhang Cui*, Yiyang Zhou*, Xinyu Yang, Shirley Wu, Linjun Zhang, James Zou, Huaxiu Yao

<div align="center"> *Equal Contribution </div> <div align="center"> <a href="https://arxiv.org/pdf/2311.03287.pdf"><img src="assets/Paper-Arxiv-orange.svg" ></a> </div>

🚨 Revealing the minds of GPT-4V(ision)! We're exploring the Limitations of Visual Language Models!

🚨 Our dataset is now available on Hugging Face, thanks to Haoqin's contribution.

Getting Started

Git clone our repository, creating a python environment and ativate it via the following command

git clone https://github.com/gzcch/Bingo.git
cd Bingo

Download the dataset from here

cd Bingo_benchmark
pip install -r requirements.txt

Convert your model response file into the format of a file like here according to the annotation file here.

python bingo_gpt_eval.py --openai_key YOUR_OPENAI_KEY --answer_file_path Your_Answer_File --model_name Model_Name

Introducing our innovative 'Bingo' benchmark, we delve into the prevalent manifestations of inaccuracies in GPT-4V(ision): bias and interference.

None

Region Bias

🌍GPT-4V(ision) favors Western images over those from other regions (e.g., East Asian, Africa) and exhibits region bias.

<div style="display: flex; justify-content: center;"> <img src="assets/regio.png" alt="None" style="width: 45%; margin-right: 5px;"> <img src="assets/region_2.jfif" alt="None" style="width: 45%; margin-left: 5px;"> </div>

OCR Bias

GPT-4V(ision) outperforms in English & French text recognition within images compared with other three languages.

<div style="display: flex; justify-content: center;"> <img src="assets/regio.png" alt="None" style="width: 45%; margin-right: 5px;"> <img src="assets/language_2.jfif" alt="Noe" style="width: 45%; margin-left: 5px;"> </div>

Factual bias

GPT-4V(ision) gets tripped by images with counterfactuals, sticking to what's 'common sense' instead of what's in the image. Like missing Saturn in a solar system photo, it still calls out Saturn.

<div style="display: flex; justify-content: center;"> <img src="assets/factual.png" alt="None" style="width: 45%; margin-right: 5px;"> <img src="assets/factual_main.png" alt="None" style="width: 45%; margin-left: 5px;"> </div>

Image-to-Image Interference

Composite images lead to confusion! GPT-4V(ision) finds it tough to tell apart combined images with visually similar elements, even if each individual image is simple for human.

<div style="display: flex; justify-content: center;"> <img src="assets/i2i.png" alt="None" style="width: 45%; margin-right: 5px;"> <img src="assets/i2i_2.jfif" alt="None" style="width: 45%; margin-left: 5px;"> </div>

Text-to-Image Interference

GPT-4V(ision) might just follow a misleading text prompt, ignoring the actual image.

<div style="display: flex; justify-content: center;"> <img src="assets/t2i.png" alt="None"> </div>

Citation

If you found this work useful, consider citing our paper as followed:

@misc{cui2023holistic,
      title={Holistic Analysis of Hallucination in GPT-4V(ision): Bias and Interference Challenges}, 
      author={Chenhang Cui and Yiyang Zhou and Xinyu Yang and Shirley Wu and Linjun Zhang and James Zou and Huaxiu Yao},
      year={2023},
      eprint={2311.03287},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}