Scaling Vision-Language Models Is Not Enough to Mitigate Bias
Large-scale evaluation of 194 VLMs shows model size correlates with overall accuracy (ρ=0.68), but not with robustness to multi-attribute bias; data quality is more critical.
Key Findings
Methodology
This large-scale empirical study analyzes 194 publicly available VLMs across 16 architectures, trained on 24 datasets, evaluating on ImageNet, CelebA, UrbanCars. Using Spearman correlation, it examines how model size, data scale, and architecture influence robustness to biases, especially multi-attribute biases. The study employs Worst-Group Accuracy (WGA) as a key metric, performing controlled experiments to disentangle effects of each factor, providing comprehensive insights into bias sensitivity across model scales and data properties.
Key Results
- Model size strongly predicts ImageNet accuracy (ρ=0.68), but its correlation diminishes significantly for single-attribute bias (ρ=0.48) and is negligible for multi-attribute bias (ρ=0.05). Data size and quality maintain more consistent relationships with WGA, with curated datasets improving robustness by up to 25%. Architectural choices like patch size and image resolution impact bias robustness contextually, with smaller patches and higher resolutions generally beneficial. Overall, increasing model size alone does not guarantee improved bias robustness, highlighting the importance of data curation.
Significance
This research challenges the common assumption that larger models are inherently more robust to biases. It emphasizes the critical role of training data quality and curation, especially for complex biases involving multiple attributes. The findings inform both academia and industry, guiding the development of fairer, more reliable multimodal systems. By revealing the limitations of scale-based improvements, it advocates for data-centric approaches to mitigate bias, contributing to more equitable AI deployment.
Technical Contribution
The study introduces the concept of bias complexity sensitivity, systematically quantifies how model performance diverges from traditional scaling laws under complex biases, and demonstrates that data quality outweighs model size in robustness. It employs rigorous statistical analysis and controlled experiments to disentangle effects, providing a comprehensive framework for evaluating bias robustness in VLMs. This work bridges the gap between performance scaling and fairness, offering new theoretical and practical insights.
Novelty
This is the first large-scale, systematic evaluation of VLM robustness to multi-attribute biases, integrating over 190 models, diverse datasets, and multiple bias scenarios. It moves beyond the conventional focus on overall accuracy, highlighting the limited impact of scale on complex biases and emphasizing data curation as a key factor. The introduction of bias complexity sensitivity provides a novel lens for understanding and improving model fairness.
Limitations
- The evaluation focuses on specific benchmarks (ImageNet, CelebA, UrbanCars), which may not capture all real-world bias scenarios. The models tested are limited to publicly available ones, possibly missing recent advances. The bias metrics primarily target visual attributes, leaving other bias types less explored. Future work should extend to more diverse biases and real-world applications.
Future Work
Future research should explore integrating bias detection and correction mechanisms into training pipelines, develop adaptive data curation strategies, and evaluate models in more complex, real-world environments. Combining active learning with bias mitigation could further enhance robustness. Additionally, extending analysis to other modalities and bias types will deepen understanding and foster fairer AI systems.
AI Executive Summary
The rapid advancement of vision-language models (VLMs) like CLIP has revolutionized multimodal AI, enabling zero-shot classification and broad applications. However, as these models are deployed at scale, concerns about their robustness to biases—especially complex, multi-attribute biases—have grown. Traditional wisdom suggests that increasing model size and training data volume leads to better performance and fairness, but recent evidence indicates this may not hold true for bias mitigation.
This study conducts a comprehensive evaluation of 194 publicly available VLMs across diverse architectures, scales, and training datasets. The models are tested on three benchmarks: ImageNet for overall recognition, CelebA for single-attribute bias (hair color vs. gender), and UrbanCars for complex multi-attribute biases involving background and co-occurring objects. The key metric, Worst-Group Accuracy (WGA), measures the model’s worst-case performance across biased subgroups.
Results reveal that while model size correlates strongly with overall accuracy (ρ=0.68), its relationship with bias robustness diminishes significantly as bias complexity increases, dropping to near zero for multi-attribute biases. Conversely, data properties—particularly dataset size and curation quality—show a more consistent and substantial impact on WGA, with curated datasets improving robustness by up to 25%. Architectural choices like patch size and image resolution influence bias performance but are highly context-dependent.
These findings challenge the prevalent focus on scale as a proxy for fairness, emphasizing the importance of high-quality, curated data in developing robust, equitable models. The research underscores that addressing bias requires a data-centric approach, integrating careful dataset curation and bias-aware training strategies. Moving forward, combining data quality improvements with advanced bias detection and correction methods will be essential to build truly fair and reliable multimodal AI systems.
Deep Analysis
Background
随着多模态学习的兴起，视觉-语言模型（VLM）如CLIP实现了零样本分类的突破，推动了内容检索、内容过滤等应用的发展。早期研究如Radford等提出的对比学习机制，强调在大规模数据和模型规模上的性能提升。近年来，偏差和公平性问题逐渐成为焦点，学界关注模型在偏差场景中的鲁棒性。已有工作多关注单一偏差或有限数据集，缺乏系统性分析。模型规模不断扩大，但偏差问题依然严重，尤其在复杂偏差场景中表现不佳。数据质量、偏差检测和修正成为研究热点，但模型规模与偏差鲁棒性关系的系统性研究仍不足，限制了公平性优化的理论基础。
Core Problem
当前多模态模型在提升整体性能的同时，偏差问题依然突出，尤其在多属性偏差场景中表现不佳。模型规模扩展未必改善偏差鲁棒性，反而可能加剧偏差依赖。如何在保证性能的同时，提高模型对复杂偏差的抗干扰能力，成为亟待解决的难题。模型在实际应用中面临多源偏差干扰，影响公平性和可靠性。缺乏系统性分析不同设计因素对偏差鲁棒性的影响，限制了模型优化策略的制定。
Innovation
本研究提出偏差复杂度敏感性概念，首次大规模系统评估模型规模、数据规模、架构参数对偏差鲁棒性的影响。通过分析16个模型家族、24个数据集，结合三项偏差基准，揭示模型在多属性偏差中的表现逐渐偏离传统规模-性能规律。创新点包括：• 采用WGA指标衡量偏差鲁棒性，• 控制模型规模与数据质量的对比实验，• 细化架构参数对偏差的影响分析，• 提出数据质量优先的模型优化策略。这些创新为多模态模型的公平性设计提供了理论基础。
Methodology
- �� 采集194个公开VLM模型，涵盖16个家族、参数规模63M至3.6B。• 训练数据包括24个不同来源，规模从13M到12.8B样本。• 评估模型在ImageNet、CelebA、UrbanCars上的表现，采用WGA指标衡量偏差鲁棒性。• 通过控制模型规模、数据规模、架构参数（patch size、图像分辨率）进行对比实验。• 统计分析模型性能与偏差复杂度的关系，验证偏差敏感性。
Experiments
采用标准零样本测试协议，ImageNet使用80个提示模板，评估top-1准确率。UrbanCars为二分类任务，检测背景和共现物体偏差，使用合成图像。CelebA为性别偏差任务，基于发色分类。模型参数包括不同模型家族、参数规模、图像分辨率和patch大小。对比不同模型在偏差场景中的表现，特别关注最大偏差组的最差组准确率（WGA）。控制模型规模与数据质量的关系，验证筛选数据对偏差鲁棒性的提升效果。
Results
模型规模在ImageNet上表现优异（ρ=0.68），但在UrbanCars偏差任务中几乎无相关（ρ=0.05），显示规模扩展对复杂偏差无效。高质量数据集能提升WGA最高达25%，架构参数如patch size和图像分辨率对偏差表现影响依赖偏差类型。控制实验显示，单纯扩大模型参数在多偏差场景中反而可能降低鲁棒性。数据规模和质量成为提升偏差鲁棒性的关键因素，强调数据筛选的重要性。
Applications
该研究为多模态模型的公平性优化提供理论指导，适用于自动驾驶、医疗影像、内容过滤等领域。通过优化数据策略和模型设计，提升模型在多偏差环境中的表现，增强系统的可靠性与公平性。未来可结合偏差检测和主动学习技术，持续改善模型鲁棒性。
Limitations & Outlook
实验主要基于公开模型和数据集，实际偏差类型更复杂多样。模型架构有限，未涵盖所有最新变体。偏差评估指标主要集中在特定任务，未覆盖所有偏差类型。未来需拓展偏差类型和场景，提升模型泛化能力。
Plain Language Accessible to non-experts
想象你在厨房里做饭，厨师（模型）需要根据食材（图片和文字）做出判断。大厨用的食谱（训练数据）越丰富，厨艺越好，但如果食谱里偏重某些调料（偏差），厨师可能只会用那种调料做菜。即使厨师变得更大、更快（模型规模扩大），如果食谱偏差没有改善，做出来的菜（模型输出）仍可能有偏。真正好的厨师，不仅要有丰富的食材，还要用多样的食谱，才能做出公平、健康的菜肴。这个比喻说明，模型的公平性和鲁棒性，关键在于食材和食谱的质量，而非单纯追求厨师的大小。
ELI14 Explained like you're 14
想象你在学校里参加比赛，大家都用不同的技巧（模型）来赢得比赛。可是，有些技巧虽然看起来厉害，但其实只是利用了比赛规则的小漏洞（偏差），而不是真正的实力。比如，有的同学会用特殊的手势（偏差特征）来作弊，但这并不能代表真正的能力。这个研究发现，变得更强（模型变大）并不能保证你不会被这些漏洞骗到，反而，使用更好的训练材料（高质量数据）能让你更公平、更稳健。就像学习要靠扎实的基础知识，而不是靠作弊技巧一样。未来，我们要用更公平的训练方法，让每个人都能真正展现自己的实力，而不是靠偏差取胜。
Abstract
Vision-Language Models (VLMs) such as CLIP are now foundational to multimodal systems, yet their robustness to spurious correlations remains poorly understood at scale. We present the first large-scale empirical study of 194 publicly available VLMs, including 16 model families, covering a wide range of model sizes, 24 training datasets, and three evaluation benchmarks, namely ImageNet (overall performance), CelebA (typical single-attribute bias), and UrbanCars (complex multi-attribute biases). Across these settings, the Spearman correlation between model scale and performance weakens as evaluation shifts from ImageNet ($ρ{=}0.68$) to single-attribute ($ρ{=}0.48$) and further to multi-attribute ($ρ{=}0.05$) bias benchmarks. In contrast, properties of the training data (size and quality) show more consistent relationships with worst-group accuracy across both bias benchmarks. Notably, curated datasets yield improvements of up to 25% over uncurated alternatives at a comparable scale. Finally, the effect of architectural choices (e.g., patch size, image resolution) is highly context-dependent, varying with the nature of the benchmark, including the type of bias and its spatial distribution within images.