Reproduce “VLBiasBench: A Comprehensive Benchmark for Evaluating Bias in Large Vision-Language Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sibo Wang, Xiangkui Cao, Jie Zhang, Zheng Yuan, Shiguang Shan, Xilin Chen, Wen Gao
VLBiasBench, a comprehensive benchmark, evaluates social biases in Large Vision-Language Models using a large-scale synthetic image dataset and multi-dimensional questioning. Its evaluation of 15 open-source and two closed-source LVLMs exposed varying degrees of bias across nine independent and two intersectional social categories.
Key insights
- Biases in LVLMs are highly varied, manifesting differently across social categories, questioning formats, and individual model architectures.
- Closed-source models generally exhibit lower bias, often due to their internal safety mechanisms that lead to more cautious or evasive responses.
- Open-source LVLMs frequently demonstrate better performance when text provides dominant evidence compared to image-dominated evidence, suggesting challenges in visual information processing.
- Fairness robustness in models degrades significantly when confronted with multi-group comparisons or ambiguous contexts, indicating susceptibility to subtle biases.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.