arXivDaily arXiv每日学术速递 周一至周五更新
​通知:联邦节假日调休:哥伦布日 / 原住民日(Columbus Day / Indigenous Peoples' Day),今日(13日)无更新。
arXiv 2510.00047cs.CVcs.AI

面向视觉-语言模型解释忠实性的解释驱动反事实测试

Explanation-Driven Counterfactual Testing for Faithfulness in Vision-Language Model Explanations

Mercedes-Benz Research & Development North America(梅赛德斯-奔驰研究与开发北美)

Sihao Ding, Santosh Vasa, Aditi Ramadwar

AI总结:

针对视觉-语言模型自然语言解释合理性与忠实性不匹配的问题,提出全自动化的解释驱动反事实测试方法,通过解析解释、生成反事实编辑、计算反事实一致性得分,验证解释的因果忠实性并提供合规审计依据。

摘要:

视觉-语言模型(Vision-Language Models, VLMs)通常会生成流畅的自然语言解释(Natural Language Explanations, NLEs),这类解释听起来很有说服力,但可能并未反映驱动预测的因果因素。这种合理性与忠实性的不匹配会带来技术和治理风险。我们提出了解释驱动反事实测试(Explanation-Driven Counterfactual Testing, EDCT),这是一种针对目标VLM的全自动化验证流程,将模型自身的解释视为可证伪的假设。给定一个图像-问题对,EDCT会:(1)获取模型的答案和NLE;(2)将NLE解析为可测试的视觉概念;(3)通过生成式图像修复生成针对性的反事实编辑;(4)借助大语言模型(LLM)分析答案和解释的变化,计算反事实一致性得分(Counterfactual Consistency Score, CCS)。在120个精心整理的OK-VQA样本和多个VLM上的测试显示,EDCT发现了显著的忠实性差距,并提供了符合监管要求的审计凭证,指明被引用的概念何时未通过因果检验。

英文摘要:

Vision-Language Models (VLMs) often produce fluent Natural Language Explanations (NLEs) that sound convincing but may not reflect the causal factors driving predictions. This mismatch of plausibility and faithfulness poses technical and governance risks. We introduce Explanation-Driven Counterfactual Testing (EDCT), a fully automated verification procedure for a target VLM that treats the model's own explanation as a falsifiable hypothesis. Given an image-question pair, EDCT: (1) obtains the model's answer and NLE, (2) parses the NLE into testable visual concepts, (3) generates targeted counterfactual edits via generative inpainting, and (4) computes a Counterfactual Consistency Score (CCS) using LLM-assisted analysis of changes in both answers and explanations. Across 120 curated OK-VQA examples and multiple VLMs, EDCT uncovers substantial faithfulness gaps and provides regulator-aligned audit artifacts indicating when cited concepts fail causal tests.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑