On the morning of April 21, Associate Professor Dong Yuxiao from the Department of Computer Science at Tsinghua University was invited by our school to deliver an academic lecture titled "Exploring the Enhancement of Agent Capabilities in Foundational Large Models" in Room 706, Building A of the Feicui Sci-Tech Complex. The event was hosted by Professor Hong Richang, Dean of the school.

A renowned scholar in artificial intelligence, Assoc. Prof. Dong Yuxiao previously worked at Facebook AI Research (FAIR) and Microsoft Research. His research focuses on foundational large models, data mining, and graph machine learning, with his academic achievements widely applied in multiple large-scale model products and systems. He was honored with the 2022 ACM SIGKDD Rising Star Award and selected as one of Tsinghua University’s "Top 10 Favorite Faculty" in 2024 (a biennial award), demonstrating excellence in both research and teaching.

In his lecture, Assoc. Prof. Dong pointed out that foundational large models exhibit strong generalization capabilities in intent perception, instruction following, and goal planning, providing a universal framework for agent research and applications. However, GUI agents based on large models still face challenges in cross-modal information fusion and cross-domain generalization. He shared his team’s cutting-edge research in three key areas: Developing a unified training and evaluation framework (AndroidLab and AgentBench) for autonomous agents, featuring a standardized operational environment that supports text input and multimodal interactions (e.g., screen-based actions like clicking and swiping), serving as a benchmark for agent development. Designing a self-evolving online curriculum reinforcement learning framework (WebRL) to address critical challenges in GUI agent training, such as task scarcity, sparse feedback, and policy drift, significantly improving performance across multiple agent benchmarks. Successful application of these models and techniques in Zhipu AI’s AutoGLM agent products, showcasing the potential of deep industry-academia collaboration.

Following the lecture, a lively Q&A session ensued, with faculty and students engaging in in-depth discussions with Prof. Dong on technical details and practical applications. Attendees praised the lecture for its forward-looking insights and real-world relevance, noting its value in inspiring new research directions.

This event not only strengthened academic ties between our university and leading institutions but also elevated our school’s research in AI. Moving forward, the school will continue to embrace an open academic philosophy, organizing more high-level exchanges to broaden perspectives and foster innovation.