Wang Zhongyuan

The official IFTC biography records Wang’s bachelor’s, master’s and doctoral degrees in computer science at Renmin University of China. It describes previous work at Microsoft Research Asia, Facebook, Meituan and Kuaishou. His earlier tutorial biography identifies research on short-text understanding and knowledge mining. At BAAI, he presented the institute’s 2026 research progress report across models, agents and computing infrastructure.[1][2][5]

Choose a useful level of abstraction

Wang’s coauthored categorization paper asks how a system can choose an appropriate conceptual label for an object. A very broad category may say too little; an overly specific one may not match the task. The method uses typicality and pointwise mutual information within a knowledge base, then compares measures and demonstrates an application to sponsored search. This connects extracted knowledge with a concrete language-understanding decision.[3]

Understand and generate several media

He also coauthored Emu3, whose training represents images, text and video as discrete token sequences. A single transformer predicts successive tokens rather than assembling separate generation and perception architectures. The report evaluates both kinds of task and describes video generation within the same formulation. Wang’s contribution is part of the named research team, distinct from his institutional leadership role.[4]