Xu Mengdi, Back from Stanford Last Year, Becomes Head Teacher of Tsinghua's Yao Class

Tsinghua University's Yao Class — widely regarded as China's flagship undergraduate AI program — has opened its 2026 cohort. According to statistics from the "Shuimu TsinghuaCent" WeChat account, the new class admitted about 90 students: roughly 66 who entered early through competition recommendation channels, plus 24 selected from Tsinghua freshmen through written exams and interviews. A significant share are national training-team members, including about 27 top informatics competitors, national-team members who represented China at international olympiads, and top gaokao scorers; one second-round admit even holds a CMO gold medal.

Standing before these elite newcomers to introduce the AI direction is a relatively new face. Xu Mengdi, an assistant professor at Tsinghua's Institute for Interdisciplinary Information Sciences (IIIS), returned from Stanford in September 2025. Less than a year into the job, she is already one of the four freshman head teachers of the Yao Class, and her orientation theme was "Building the Next Generation of Intelligence."

Xu is neither a Yao Class graduate nor a traditional computer science student. She earned her bachelor's degree in vehicle engineering from Tsinghua in 2017, then studied robotics at Johns Hopkins University and machine learning at Carnegie Mellon University, receiving a PhD in mechanical engineering from CMU in 2024. Under CMU Safe AI Lab head Ding Zhao, she researched robot learning, reinforcement learning, and AI safety.

One of her signature doctoral works is RoboTool, developed with researchers from CMU and Google DeepMind. Instead of specifying step by step which tools a robot should use and which actions to take, the system receives a task goal plus scene, object, and robot-capability information, then analyzes constraints, chooses tools, plans steps, and generates executable code. In demonstrations, a robotic arm that could not reach a distant milk carton used a hammer as a hook to pull it closer, and a quadruped robot unable to cross a gap between sofas built a path from nearby objects.

After her PhD, Xu did postdoctoral research at Stanford's Vision and Learning Lab with Jiajun Wu and Fei-Fei Li — a major hub for embodied AI. A year later, she chose to return to her undergraduate alma mater, joining Tsinghua's IIIS to build her own research group.

More notable is her stance on generalist robotics. In her doctoral thesis, "Building Adaptable Generalist Robots," she directly challenges the scaling-law assumption that bigger models and more data naturally produce generalization. The real world's task space is nearly infinite, so no training set can cover everything — robots need the ability to adapt quickly on site from small amounts of new information.

The IIIS, founded by Andrew Yao in 2011, is the home institution of the Yao Class, with research spanning theoretical computer science, AI, and quantum information. A young scholar who started in vehicle engineering and later turned to robot learning now stands on the Yao Class podium, adding a fresh embodied-intelligence perspective to an institute long known for theory and large models.

The question worth watching is what this head teacher, with her "Building the Next Generation of Intelligence" theme, brings into the Yao Class classroom — and what her team delivers next in robot learning and embodied AI.