一句话：提出以自动化测试刻画与运行时监控相结合的范式，为不透明 AI 模型赋能的关键系统提供可信保障。
将不透明 AI 模型快速集成到关键软件基础设施中，给系统可靠性与可信性带来了根本性挑战。本文以“通过系统性交互获得洞察来解码复杂性”为理念，围绕两个研究问题组织工作：RQ1 测试与刻画，关注自动化输入生成与边界探测，以揭示模型隐藏行为；RQ2 增强与保障，利用挖掘出的行为模式来构建主动防御与运行时监控。作者在多种 AI 模型和多样化任务上进行了大量成功尝试，并进一步提出未来议程，旨在将该范式扩展到更广泛的 AI 生命周期与更复杂的 AI 系统中。
The rapid integration of opaque AI models into critical software infrastructure creates a fundamental challenge for reliability and trustworthiness. This paper addresses this dilemma through the philosophy of Decoding Complexity by Gaining Insight through Systematic Interaction. We present a body of work organized around two research questions: Testing & Characterization (RQ1), which focuses on automated input generation and boundary probing to reveal hidden behaviors, and Enhancement & Assurance (RQ2), which leverages mined behavioral patterns to engineer active defenses and runtime monitoring. Having made numerous successful attempts across a broad spectrum of AI models and diverse tasks, we outline a future agenda to scale this paradigm to the broader AI lifecycle and more complex AI systems.