Melange: A Synthetic Visual Bias Benchmark for Multi-modal Large Language Models
作者: Razieh Rezaei, Ehsaneddin Asgary, Masoud Jalili Sabet
分类: cs.CV
发布日期: 2026-10-21
💡 一句话要点
提出Melange基准以评估多模态大语言模型的社会偏见
🎯 匹配领域: 支柱九:具身大模型 (Embodied Foundation Models)
关键词: 多模态大语言模型 社会偏见 合成视觉基准 偏见分析 视觉问答 人工智能伦理 公平性评估
📋 核心要点
- 现有的偏见分析方法在多模态大语言模型中存在覆盖范围不足和视觉线索不明确的问题。
- Melange基准通过合成视觉场景,确保目标属性在图像中清晰可见,从而全面评估MLLMs的社会偏见。
- 实验结果表明,切换输入模态从文本到视觉平均降低公平性0.291,且这一差距因模型家族而异,而非仅由规模决定。
📝 摘要(中文)
多模态大语言模型(MLLMs)在广泛应用中,其偏见问题尚未得到充分研究,可能在实际应用中造成社会危害。现有的偏见分析方法要么只涵盖少数偏见类别,要么未能确保比较组在图像中有明确的视觉线索。为此,本文提出了Melange,一个涵盖七个类别的综合性合成视觉基准,用于评估MLLMs中的社会偏见。该基准基于广泛使用的语言-only BBQ数据集,通过控制生成的视觉场景扩展其构建的上下文,以确保目标属性在视觉上清晰可见。生成的图像经过自动筛选和人工检查两个阶段的验证,确保可靠性。通过在文本、图像和图像与上下文三种匹配评估设置下系统评估23个开源MLLMs,Melange有效隔离了输入模态的公平性影响。
🔬 方法详解
问题定义:本文旨在解决多模态大语言模型(MLLMs)中偏见分析不足的问题,现有方法未能全面覆盖偏见类别且缺乏明确的视觉线索。
核心思路:Melange基准通过合成视觉场景,确保目标属性在图像中清晰可见,从而提供一个全面的评估工具,以揭示MLLMs的社会偏见。
技术框架:整体架构包括两个主要阶段:首先是通过视觉问答(VQA)进行自动筛选,随后由人工注释者进行手动检查,确保生成图像的质量和准确性。
关键创新:Melange的创新在于其合成视觉基准的构建,确保了目标属性在视觉上清晰可见,这与现有方法的模糊性形成鲜明对比。
关键设计:在生成图像时,采用了控制生成的策略,确保每个类别的偏见在视觉上得到充分体现,同时在验证阶段使用了自动化和人工双重检查机制。
🖼️ 关键图片
📊 实验亮点
实验结果显示,切换输入模态从文本到视觉平均降低公平性0.291,且这一差距在不同模型家族中有所不同。尽管模型在识别图像中的属性时准确率达到90-95%,但在偏见回答的选择上,约75%的情况下仍倾向于人类刻板印象的反应。
🎯 应用场景
该研究的潜在应用领域包括社会科学研究、人工智能伦理和多模态系统的开发。通过提供一个标准化的偏见评估工具,Melange可以帮助研究人员和开发者识别和减轻多模态模型中的社会偏见,从而提高其在实际应用中的公平性和可靠性。
📄 摘要(原文)
Multi-modal Large Language Models (MLLMs) have seen broad adoption, but their biases are still far less examined than those of standard LLMs, creating possible social harms when deployed in real-world settings. Prior efforts to extend bias analyses from LLMs into the multi-modal space either cover only a few bias categories or do not ensure that the groups being compared are clearly addressed by visual cues in the images. To overcome these limitations, we introduce Melange, a wide-ranging synthetic visual benchmark covering seven categories for assessing social biases in MLLMs. It is grounded in the widely used language-only BBQ dataset (Bias Benchmark for Question Answering) from BigBench, carefully extending its constructed contexts via controlled generation of visual scenarios so the target attributes appear visually and clearly. To maximize reliability, our generated images pass through two stages of validation: automated screening using visual question answering (VQA) and subsequent manual inspection by human annotators. By systematically evaluating 23 open-source MLLMs across three matched evaluation settings---text-only, image-only, and image-with-context---Melange isolates the fairness effect of the input modality. Switching from text to vision reduces fairness by 0.291 on average, a gap that varies by model family rather than by scale alone. A perception control confirms that this gap is not an artifact of visual comprehension: models identify the depicted attributes at 90--95\% accuracy yet still fail to abstain. When models do commit to a biased answer, polarity-aware stereotype alignment shows they choose the human-stereotypical response roughly three-quarters of the time.