Research map / Vision-language and multimodal / Multimodal LLMs
VLBiasBench: A Comprehensive Benchmark for Evaluating Bias in Large Vision-Language Model
Sibo Wang, Xiangkui Cao, Jie Zhang, Zheng Yuan, Shiguang Shan, Xilin Chen, Wen Gao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase LLaVA-1.5 size from 7b to 13b | LLaVA-1.5 · VLBiasBench open-ended race | Range VADER: 0.065 -> 0.044 | Improved |
| increase InstructBlip-flan-t5 language model from xl to xxl | InstructBlip · VLBiasBench open-ended race | Range VADER: 0.225 -> 0.014 | Improved |
| move from text-dominant (Base) to image-dominant (Scene) evidence | LLaVA-1.5-13b · VLBiasBench close-ended | ACC: 0.578 -> 0.375 | Hurt |
| move from text-dominant (Base) to image-dominant (Scene) evidence | Gemini · VLBiasBench close-ended | ACC: 0.672 -> 0.876 | Improved |
| add a second compared group via text-induced questions (Base to Text) | InstructBlip-flan-t5-xxl · VLBiasBench close-ended | ACC: 0.772 -> 0.675 | Hurt |
| add a second compared group via text-induced questions (Base to Text) | GPT-4o · VLBiasBench close-ended | ACC: 0.882 -> 0.773 | Hurt |
| add a second compared group via text-induced questions (Scene to Scene Text) | InstructBlip-flan-t5-xxl · VLBiasBench close-ended | ACC: 0.616 -> 0.433 | Hurt |
| use ambiguous instead of disambiguated contexts | LLaVA-1.5-7b · VLBiasBench close-ended Base | ACC: 0.763 -> 0.001 | Hurt |
| use ambiguous instead of disambiguated contexts | Gemini · VLBiasBench close-ended (All) | ACC: 0.553 -> 0.988 | Improved |
| apply sentiment thresholds to filter neutral responses (ablation) | Range VADER metric · VLBiasBench open-ended | rank stability | No clear effect |
| reverse the close-ended questions to test selection bias (ablation) | close-ended benchmark · VLBiasBench close-ended | model rank: 7 -> 6 | No clear effect |
| score sentiment with GPT instead of VADER (ablation) | Range VADER metric · VLBiasBench open-ended race | Pearson correlation of ranks: 0.978 | No clear effect |
Closest papers by meaning
- VHELM: A Holistic Evaluation of Vision Language Models (NeurIPS 2024 · similarity 0.91)
- Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation (ICML 2026 · similarity 0.90)
- Are We on the Right Way for Evaluating Large Vision-Language Models? (NeurIPS 2024 · similarity 0.90)
- SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples (CVPR 2024 · similarity 0.90)
- On Fairness of Unified Multimodal Large Language Model for Image Generation (NeurIPS 2025 · similarity 0.90)
- Joint Vision-Language Social Bias Removal for CLIP (CVPR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
