一句话：研究 LLM 智能体如何解释与推理软件视觉上下文，并设计更有效利用图像线索的多模态智能体。
该博士研究关注软件工程中的视觉制品（如截图、示意图等）在基于大语言模型（LLM）的智能体中的利用问题。视觉信息对软件问题求解至关重要，但现有多模态语言模型在有此类制品时提升有限，图像常被低效利用或误读，暴露出多模态信息与软件工程任务整合的薄弱环节。研究将实证分析 LLM 智能体如何解释和推理软件特定的视觉上下文，并与人类理解进行对比；在此基础上设计并评估能更有效融入视觉线索进行推理与行动的多模态智能体。最终目标是在真实软件工程场景中提升多模态问题求解能力。
Visual artifacts play a critical role in software engineering problem solving, yet current multimodal language models show limited improvement when such artifacts are available. Recent evidence suggests that images are often underutilized or misinterpreted, revealing fundamental weaknesses in how multimodal information is integrated into software engineering tasks. This doctoral research investigates how LLM-based agents interpret and reason over software-specific visual context, and how these interpretations differ from human understanding. Beyond empirical analysis, the work designs and evaluates multimodal agents that more effectively incorporate visual cues during reasoning and action. The goal is to develop agents that improve multimodal problem solving in real-world software engineering settings.