🤖 AI

VLBiasBench: A Comprehensive Benchmark for Evaluating Bias in Large Vision-Language Model

This paper introduces VLBiasBench, a comprehensive benchmark featuring a large-scale dataset of over 128,000 samples across 11 bias categories and diverse question formats, designed to rigorously evaluate and reveal social biases in both open-source and closed-source Large Vision-Language Models.

Original authors: Sibo Wang, Xiangkui Cao, Jie Zhang, Zheng Yuan, Shiguang Shan, Xilin Chen, Wen Gao

Published 2026-04-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sibo Wang, Xiangkui Cao, Jie Zhang, Zheng Yuan, Shiguang Shan, Xilin Chen, Wen Gao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a group of incredibly smart, super-fast robots. These robots can see pictures and read text, and they can talk to you about what they see. They are like the ultimate "see-and-say" assistants. But, just like humans, these robots have learned from a massive library of books and images created by people. And unfortunately, that library has some old-fashioned, unfair ideas baked into it.

This paper introduces VLBiasBench, which is essentially a giant, high-tech "spot-the-bias" test designed to see if these robots are being unfair.

Here is the breakdown of what they did, using some everyday analogies:

1. The Problem: The "Broken Library"

Think of these AI models as students who studied in a giant library. If that library has more books about "men being doctors" and "women being nurses," or "certain races being criminals," the students (the AI) will learn those patterns as facts. When you ask them a question, they might give you an answer that sounds smart but is actually full of stereotypes.

The problem is, we didn't have a good way to test all these robots on all these different types of unfairness (like age, race, religion, disability, etc.) at the same time. Previous tests were too small or only looked at one thing.

2. The Solution: The "Synthetic Playground"

To fix this, the researchers built a brand new testing ground called VLBiasBench.

The Magic Paintbrush (Image Generation): Instead of using real photos of real people (which could accidentally include people who already "studied" the AI's training data), they used a magic paintbrush called Stable Diffusion. They asked this brush to paint thousands of pictures of people with specific traits (e.g., "an elderly Asian man," "a young Black woman"). This ensures the test is fresh and fair.
The Question Master (The Prompts): They didn't just show the pictures; they asked the robots tricky questions.
- Open-Ended Questions: "Tell me a story about this person." (This is like asking a student to write an essay. If the student writes a story where the "Black man" is a criminal and the "White man" is a CEO, you know they have a bias.)
- Multiple-Choice Questions: "Is this person suitable for this job?" with options like "Yes," "No," or "I don't know." (This is like a standardized test to see if the robot guesses based on stereotypes or facts.)

3. The Test: The "Fairness Olympics"

They put 17 different AI models (both the free, open-source ones and the fancy, paid ones like GPT-4 and Gemini) through this gauntlet.

The Scorecard: They didn't just count right or wrong answers. They used a "sentiment meter" (a tool that measures if the words used are happy, sad, angry, or neutral).
- If a robot describes a woman as "emotional and weak" but a man as "strong and leader-like" for the exact same picture, the sentiment meter goes off, and the robot gets a bad score.
- They also checked if the robot was "overconfident." Sometimes, when there isn't enough information to answer, a biased robot will guess anyway based on stereotypes, while a fair robot will say, "I don't know."

4. The Results: Who Passed and Who Failed?

The results were a mix of good news and bad news:

The "Old Guard" (Older/Open-Source Models): Many of the earlier models were like students who hadn't been corrected yet. They often fell into stereotypes. For example, one model consistently assumed an older man couldn't use a smartphone, or that a woman in a certain outfit was a thief.
The "New Guard" (Advanced Closed-Source Models): The big, expensive models (like GPT-4o and Gemini) generally did much better. They were like the students who had been taught better manners. They were less likely to make unfair assumptions.
The Surprise: Even the "good" students sometimes stumbled when the questions got really tricky or when two types of bias mixed together (like race and gender).

5. Why This Matters

Think of VLBiasBench as a "report card" for the future of AI.

If we don't test these robots, they might start making hiring decisions, writing news stories, or giving medical advice based on unfair stereotypes. This paper gives us the tools to:

Catch the bias: See exactly where the robot is being unfair.
Fix the robot: Use these test results to teach the AI better.
Build a fairer future: Ensure that when AI helps us, it treats everyone with the same respect, regardless of what they look like or where they come from.

In short: The researchers built a massive, fair, and tricky test to see if our smartest robots are secretly holding onto old prejudices. They found that while some robots are improving, we still have a long way to go to make sure they are truly fair for everyone.

1. Problem Statement

Large Vision-Language Models (LVLMs) have achieved remarkable progress in general artificial intelligence but inherit and amplify social biases from their pre-training data. These biases manifest as unfair associations based on sensitive attributes such as race, gender, religion, and socioeconomic status.

Limitations of Existing Benchmarks: Current evaluation tools suffer from three main deficiencies:
1. Limited Scale: Existing datasets are often small or lack diversity.
2. Single Questioning Format: Most benchmarks rely solely on either open-ended generation or multiple-choice questions, failing to capture bias from different interaction modes.
3. Narrow Bias Sources: Many benchmarks focus on early vision-language models (e.g., CLIP) rather than advanced, large-scale LVLMs (e.g., LLaVA, Gemini). Furthermore, using real-world images risks data leakage (test samples appearing in training sets) and lacks controllability over specific bias attributes.

2. Methodology

The authors propose VLBiasBench, a comprehensive benchmark designed to evaluate biases across 11 categories (9 independent and 2 intersectional) using synthetic data to ensure controllability and prevent data leakage.

A. Dataset Construction

Image Generation: The authors utilized Stable Diffusion XL (SDXL) to generate 46,848 high-quality synthetic images.
- Prompt Engineering: A multi-stage prompt library was constructed using a combination of bias-relevant corpora (from BOLD and BBQ datasets) and automatic generation via GPT-4. Prompts included specific style and quality control words to ensure image fidelity.
- Filtering: A rigorous three-stage filtering process was applied:
  1. CLIP Semantic Filtering: Ensured image-text alignment (cosine similarity > 0.7).
  2. NSFW Filtering: Removed unsafe or offensive content using safety checkers and NudeNet.
  3. Manual Filtering: Human experts removed images with structural artifacts (e.g., fused limbs) to ensure high quality.
Question Construction: The dataset comprises 128,342 samples divided into two evaluation modes:
1. Open-Ended Evaluation: Designed to induce bias through context completion tasks (e.g., "Write a story based on the person's race").
2. Close-Ended Evaluation: Multiple-choice questions (Yes/No/Unknown) designed to test reasoning under specific constraints.
Evaluation Framework (Close-Ended): The close-ended dataset is split into four subsets to isolate bias sources:
- Base: Text-dominant evidence; images show only protected attributes (e.g., race) without context.
- Scene: Image-dominant evidence; images contain the decisive visual clues, while text is minimal.
- Text & Scene Text: Enhanced versions of Base and Scene that introduce text-induced comparative questions (e.g., "Compared with [Group A], is [Group B]...?") to test fairness in multi-group scenarios.

B. Bias Categories

The benchmark covers:

Independent Categories: Age, Disability, Gender, Nationality, Physical Appearance, Race, Religion, Profession, Socioeconomic Status (SES).
Intersectional Categories: Race $\times$ Gender, Race $\times$ SES.

C. Evaluation Metrics

Open-Ended:
- Sentiment Analysis: Uses VADER to calculate sentiment scores (range -1 to 1). Bias is measured by the Range VADER (difference between max and min average sentiment scores across subgroups). A larger range indicates higher bias.
- Gender Polarity: Measures the similarity of word embeddings to gender vectors (she/he) to detect gendered associations in profession-related responses.
Close-Ended:
- Accuracy: Measures the proportion of correct predictions.
- Robustness ( $\Delta$ ): Calculates the accuracy drop when moving from single-group (Base/Scene) to multi-group (Text/Scene Text) scenarios.

3. Key Contributions

VLBiasBench Benchmark: The first large-scale, synthetic-data-driven benchmark specifically for LVLM bias, covering 11 bias categories with 128k+ samples.
Multi-Dimensional Evaluation: Introduces a dual-mode framework (Open-ended and Close-Ended) and a four-subset structure (Base, Scene, Text, Scene Text) to comprehensively assess how models handle bias across different modalities and comparison contexts.
Comprehensive Empirical Study: Evaluates 15 open-source models (e.g., LLaVA, MiniGPT, InstructBlip) and 2 closed-source models (GPT-4o, Gemini), providing the first large-scale comparison of bias in state-of-the-art LVLMs.
Validation of Synthetic Data: Demonstrates through ablation studies (comparing with FairFace and using SD3.5/FLUX.1) that synthetic images effectively capture bias trends comparable to real-world data while avoiding data leakage.

4. Experimental Results

The evaluation of 17 models yielded several critical insights:

Model Performance Variance:
- Open-Ended: Shikra-7b exhibited the most severe biases across all categories. InstructBlip-flan-t5-xl showed high bias in race, while InstructBlip-vicuna-13b was highly biased regarding gender in professions.
- Close-Ended: InstructBlip-flan-t5-xxl and GPT-4o generally demonstrated the highest fairness (highest accuracy).
- Closed-Source vs. Open-Source: Closed-source models (GPT-4o, Gemini) consistently showed lower bias levels, likely due to strict safety mechanisms and refusal to answer subjective queries.
Modality Influence: Open-source models performed significantly better on Text-dominant (Base) questions than Image-dominant (Scene) questions, indicating a reliance on textual cues and difficulty in extracting unbiased visual evidence. Closed-source models showed more robustness across both modalities.
Intersectional Bias: Models struggled significantly when bias categories intersected (e.g., Race $\times$ SES). For instance, Gemini's performance dropped sharply in intersectional scenarios compared to single-category scenarios.
Parameter Scale: Generally, larger parameter models (e.g., LLaVA-13b vs. 7b) exhibited slightly lower bias, though this trend was not universal (e.g., InstructBlip variants showed inconsistent behavior based on the underlying LLM size).
Text-Induced Bias: Introducing comparative text (Text/Scene Text subsets) often degraded the fairness of models that performed well in single-group settings, revealing that many models lack robustness when explicitly asked to compare sensitive groups.

5. Significance

Standardization: VLBiasBench establishes a rigorous, reproducible standard for evaluating fairness in LVLMs, addressing the gap left by previous benchmarks that focused on older models or lacked scale.
Safety and Ethics: By identifying specific bias patterns (e.g., associating certain races with crime or specific genders with low-status professions), the benchmark provides actionable insights for developers to improve model alignment and safety.
Methodological Advancement: The use of synthetic data generation with strict filtering proves to be a viable, scalable, and cost-effective alternative to real-world datasets for bias research, solving the data leakage problem inherent in using public image datasets.
Future Direction: The findings highlight that while closed-source models are currently more robust, open-source models need significant improvement in visual reasoning and intersectional fairness to be deployed safely in real-world applications.

The benchmark and code are publicly available at https://github.com/Xiangkui-Cao/VLBiasBench.