一句话：将无障碍评估重构为自动化可达性问题，用 LLM 智能体对比测试发现视障用户遭遇的动态缺陷。
现有自动化可访问性扫描工具主要基于静态规则，难以捕捉视障用户在实际操作流程中遭遇的动态、工作流级障碍。本文提出一个 LLM 驱动的测试框架，把可访问性评估重新表述为自动化的可达性问题：框架将视觉化的“正常用户智能体”（用于提取任务指引）与由 TalkBack 驱动的“盲用户智能体”相耦合，后者配备贴近真实操作的导航启发式，包括近似重定位（approximate relocation）与邻域浏览（vicinity skimming）。该流水线在非视觉约束下系统性地评估任务可达性，通过两类智能体的对比暴露障碍。作者在 15 个真实 Android 应用上进行了评估，有效发现四类动态可访问性缺陷：非交互控件、导航陷阱、不当焦点顺序与导航死区。
Despite advancements in automated software testing, existing accessibility scanners struggle to detect dynamic, workflow-level barriers faced by visually impaired users. We present an LLM-driven testing framework that reformulates accessibility evaluation into an automated reachability problem. By coupling a visual Normal User Agent to extract task guidance with a TalkBack-driven Blind User Agent equipped with realistic navigation heuristics (approximate relocation and vicinity skimming), our pipeline systematically evaluates task reachability under non-visual constraints. Evaluated across 15 real-world Android applications, our framework effectively uncovers four types of dynamic accessibility defects: non-interactive control, navigation trap, improper focus ordering, and navigation dead zone.