发现视觉语言模型在面对视觉矛盾时容易误信错误描述。
Pathological Truth Bias in Vision-Language Models
- 设计新评测方法MATS，检测模型对视觉矛盾陈述的拒绝能力
- 生成式模型SCS低至0.12，错误认同率IAR高达0.78，对比模型更可靠
- 定位失败位置，为修复提供可操作路径，适合模型安全研究者
视觉语言模型（VLMs）发展迅速，但标准基准测试可能掩盖系统性缺陷，削弱真实场景信任度。本文提出MATS（多模态真实性审计），一种紧凑的行为审计方法，用于衡量模型是否拒绝与视觉内容矛盾的陈述，并定义了空间一致性得分（SCS）和错误一致率（IAR）两个指标。指令微调的生成式VLM（LLaVA 1.5、QwenVLchat）表现出极低的SCS（0.12）和高的IAR（0.78），而对比编码器（CLIP、SigLIP）则更为稳健。通过激活块替换，因果定位发现生成式模型的故障位于中间到晚期交叉注意力层，对比模型则集中在池化投影组件，提示了具体的修复方向。
原文摘要 · Abstract (English)
Vision Language Models (VLMs) are improving quickly, but standard benchmarks can hide systematic failures that reduce real world trust. We introduce MATS (Multimodal Audit for Truthful Spatialization), a compact behavioral audit that measures whether models reject visually contradicted statements, and two metrics Spatial Consistency Score (SCS) and Incorrect Agreement Rate (IAR). Instruction tuned generative VLMs (LLaVA 1.5, QwenVLchat) exhibit very low SCS and high IAR, while contrastive encoders (CLIP, SigLIP) are far more robust. Activation patching causally localizes failure loci (mid to late cross attention for generative models, pooled projection components for contrastive models) and suggests concrete repair paths.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。