Abstract
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.
Abstract (translated)
视觉-语言模型（VLM）常常会产生流利的自然语言解释（NLE），这些解释听起来非常可信，但实际上可能并不反映预测背后的原因。这种合理性和忠实性之间的不匹配带来了技术和治理上的风险。我们介绍了一种称为基于解释的反事实测试（EDCT）的完全自动化验证程序，该程序以目标VLM生成的自身解释作为可证伪假设进行处理。对于给定的一组图像和问题： 1. EDCT 首先获取模型的回答和 NLE； 2. 接着将NLE解析成可以测试的视觉概念； 3. 通过生成式修复技术（generative inpainting）来创建有针对性的反事实修改； 4. 最后，使用LLM辅助分析答案和解释的变化来计算反事实一致性得分（CCS）。 通过对120个经过挑选的OK-VQA示例以及多个VLM进行测试，EDCT揭示了显著的忠实性差距，并提供了符合监管要求的审计结果，指示哪些引用的概念未能通过因果关系检验。
URL
https://arxiv.org/abs/2510.00047
