From 15 Nodes to 389: Chinese Team Says Shanghai AI Lab Intern-S2 Helped Prove the 30-Year-Old Ghys Conjecture

A Chinese research team announced this week that it has fully proved the three-dimensional Ghys conjecture, an open problem in complex dynamics that has stood since French mathematician Etienne Ghys posed it in 1995 — and that the heavy lifting of derivation was done in collaboration with a domestically developed AI model. If the proof holds up to community scrutiny, it would be the first time a Chinese-built large model has been credited with helping crack a problem of this standing.

The work is a collaboration between Xu Disheng, a professor at Greater Bay University and a visiting scholar at the Shanghai Artificial Intelligence Laboratory, and his student Zhang Jiesong, using the lab's Intern-S2 science-focused multimodal model. The conjecture concerns holomorphic Anosov systems on compact complex manifolds — chaotic dynamical systems — and asks whether they can be brought into correspondence with rigid, algebraically tractable models. Its real-number counterpart remains open worldwide, which is why the problem has served as a benchmark for mathematical reasoning ability.

The collaboration ran as a tight human-AI loop rather than a single prompt. Intern-S2's first pass on the problem produced a 400-page derivation report on its own. Over the course of the project the model completed seven successive batches of derivation, with the working problem graph expanding from 15 nodes to 389, outputting what the team says totals more than a thousand pages of rigorous derivation and dozens of new lines of argument. The humans' job was direction: judging which avenues were worth pursuing, feeding back corrections, and steering the model toward productive territory.

Crucially, the team did not ask anyone to take the derivation on faith. A Peking University group led by Liu Ruochuan, an academician of the Chinese Academy of Sciences, and professor Dong Bin independently verified the proof using their SHEAF framework, running formal machine verification of all core theorems and producing hundreds of thousands of lines of verification code. The paper has been posted to arXiv as preprint 2609.37695 and the formalization code is open-sourced on GitHub, so the check is reproducible by anyone. The results are team-announced and the preprint is not yet journal peer-reviewed, but the formal verification means the core claims are machine-checked rather than self-attested.

There is a telling footnote about the competition. Peking University professor Xie Junyi, a 2026 International Congress of Mathematicians speaker, reportedly tried GPT-6 on the same conjecture before the Chinese result was published and did not find a proof. The project's origin story is equally data-driven: Shanghai AI Lab's earlier Intern-S1 scored 102 points in the 2025 Chinese Mathematical Olympiad's first AI test, topping the participating models and clearing the 78-point human gold-medal line, which convinced the lab to push from competition mathematics toward research-level problems.

The announcement lands in a crowded field. Google DeepMind's AlphaProof line has been racking up solved Erdos problems, and OpenAI's mathematics push has drawn public pushback from mathematicians over rigor and reproducibility. What distinguishes the Ghys result is the verification-first posture: the claim is not that a model produced an impressive-looking proof, but that a second team formalized every core theorem in machine-checkable code before anyone was asked to believe it.

The template matters more than the theorem. The team describes a loop of system exploration, expert judgment and feedback correction that they argue can be ported to any field where the search space is huge and experts are scarce. Whether that generalizes beyond one conjecture in complex dynamics is the next question — but as a demonstration that AI-assisted research can be made formally checkable, this is a serious data point.