GeoPT Lead Author Wu Haixu Found a World-Model Startup Valued in the Hundreds of Millions

The world-model race has a new academic entrant. Wu Haixu, who won Tsinghua University's top undergraduate scholarship in 2022, has quietly started a company in the space this year called 物原科技 (Wuyuan), according to reporting by Chinese tech outlet QbitAI.

The move is not a sudden pivot. Wu has spent years on physics and dynamics modeling, contributing to work such as Autoformer, PredRNN, Corrformer and Transolver. In February he was lead author on GeoPT, a pretraining approach that lifts static geometry into dynamic space through synthetic dynamics, letting a model acquire physical intuition by learning how particle trajectories evolve on unlabeled data. Co-authors included Tsinghua's Long Mingsheng and MIT's He Kaiming.

The financing came together quickly. QbitAI reports that the startup has already closed several rounds and reached a valuation in the hundreds of millions of dollars, with Ant Group, Meituan, Dachen Caizhi and Xinglian among the investors.

The product direction, according to people close to the company, is a world model with physical modeling built in, one that extends the GeoPT research line to some degree. On that description the goal is not convincing video but a computable, rollable-out representation of how objects move and interact.

Wu's academic path explains the attention. He entered Tsinghua's School of Software in 2016, added a second degree in pure and applied mathematics, spent a summer at Harvard's Visual Computing Group, completed his PhD under Long Mingsheng, and since 2025 has been a postdoc at MIT CSAIL with Wojciech Matusik. His Google Scholar citation count now stands at 29,611.

World models are increasingly treated as the missing layer between foundation models and reliable embodied systems, and academic teams are rushing in. Tsinghua's special scholarship has already produced founders such as Yang Zhilin of Moonshot AI and Cao Yue of Sand.ai, and Wu's entry suggests the field still holds its pull on frontier researchers.

Much remains undisclosed. The public record stops at multiple rounds and a valuation in the hundreds of millions of dollars; round sizes, leads, headcount and the eventual form of the product are all unknown. The question worth tracking is whether GeoPT-style physics pretraining survives contact with real robots, factories and lab benches, and what deployment scenarios Ant Group and Meituan have in mind as investors.