Chinese Scientists Use Shusheng-S2 to Prove Ghys Conjecture.

Researchers from the Shanghai AI Laboratory, in collaboration with academic institutions, have successfully proven the three-dimensional Ghys conjecture, a longstanding open problem in dynamical systems that has challenged mathematicians for over three decades. This achievement marks the first instance of Chinese scientists solving a world-class mathematical conjecture using a domestically developed large language model. The breakthrough was spearheaded by Professor Xu Disheng of Guangdong University of Technology, serving as a visiting scholar at the Shanghai AI Laboratory, alongside Dr. Zhang Jiesong. The team leveraged the Shusheng-S2 multimodal large model, which underwent rigorous training on research-grade mathematical corpora and implemented advanced asynchronous long-range reasoning reinforcement learning. Central to the methodology was the Harness system, a dedicated infrastructure enabling large-scale recursive deduction and high-flexibility human-machine interaction. From August 2026, Shusheng-S2 initiated multi-directional exploratory reasoning, generating thousands of pages of preliminary derivations and dozens of novel mathematical approaches. The AI system extensive computational search was strategically guided by human experts. Professors Xu and Zhang continuously evaluated the model outputs, identifying critical breakthrough points and calibrating the system directional search through iterative feedback. This collaborative loop condensed over seven major reasoning cycles, expanding the problem graph from 15 to 389 nodes and ultimately synthesizing a complete proof chain. The resulting work, which international experts have noted for exceeding anticipated analytical depth, has been published on arXiv. To ensure mathematical rigor, a team led by Professor Dong Bin and Academician Liu Ruochuan at Peking University conducted a comprehensive formal verification using the Lean proof assistant and their proprietary SHEAF framework. The verification process, completed in approximately five days with peak parallel computing across 100 intelligent agents, generated over 550,000 lines of formal code. This step confirmed the logical consistency of all four primary theorems and established foundational formalization infrastructure for complex dynamical systems previously lacking computer-verified frameworks. The successful resolution of the Ghys conjecture demonstrates the practical efficacy of a closed-loop scientific discovery paradigm: autonomous system exploration, expert intuition filtering, and continuous feedback correction. This methodology significantly reduces the trial-and-error burden in high-complexity research domains. The Shanghai AI Laboratory intends to encapsulate the validated capabilities into the reusable Shusheng-Duanyan scientific discovery platform, fostering a self-evolving cycle of model refinement and cross-disciplinary application. By bridging advanced AI inference with specialized domain expertise, the breakthrough establishes a scalable template for tackling foundational scientific challenges, signaling a transformative shift toward AI-empowered original innovation in mathematics and beyond.