Jiuzhang Yunji Enters Tao's Math Challenge After Letter From 28 Fields Medalists

On 11 September 2026, twenty-eight Fields Medal winners signed a single statement, "Serious Misalignment of AI in Mathematics." Exactly a month later, on 13 October, China's Jiuzhang Yunji (九章云极) announced it would take part in Terence Tao's math challenge and set out its position: reinforcement learning should become the new base infrastructure for AI for Science. The timing looks intriguing, but the available excerpt of the source material says nothing about the format of the competition or the company's results.

What the mathematicians wrote

The letter is unusually sharp for an academic setting. The signatories argue that AI companies use math problem solving as a benchmark of their own capabilities, and they consider this practice "harmful to mathematical science and the mathematical community." Their main thesis, as far as the opening of the material shows, is that solving a problem is only a means and a proxy metric. Mathematics itself works differently: it is the construction of understanding, theories and language. The argument is cut off in the available excerpt, so retelling its details would be guesswork.

For the industry, this is an uncomfortable signal. Olympiad-level and research-level problems have served as a showcase for AI labs for the past couple of years. When 28 people holding the profession's highest award say the showcase is misleading, it becomes harder to answer with numbers.

Jiuzhang Yunji's response

The company chose not to defend itself but to enter a competition run by one of the most respected mathematicians of our time. The slogan in the headline of the source material reads: make reinforcement learning the new infrastructure of AI4S.

The logic is easy to follow, although this is our interpretation, not a quote from the company. Mathematics has something most scientific fields lack: verifiability. If a proof can be checked automatically, for example in a formal system such as Lean, it yields a clear reward signal for RL. After that, everything comes down to engineering: orchestrating thousands of parallel runs, scheduling GPUs, storing trajectories, and building environments with verifiers. Judging by its positioning, Jiuzhang Yunji sells exactly this layer: not a standalone model, but a platform on which such experiments can be run.

What is known at this point:

• Date of the letter: 11 September 2026, with 28 Fields medalists as signatories.
• Title of the statement: "Artificial Intelligence in Mathematics: A Serious Misalignment."
• Key objection: using problems as a benchmark harms mathematics and the community.
• Participant in Tao's challenge: Jiuzhang Yunji, announced on 13 October 2026.
• Stated direction: RL as AI for Science infrastructure.
• Not disclosed: the rules, the number of problems solved, the amount of compute, the verification method.

What to watch next

So far this is a positioning move, not a result. Several signs are worth using to judge it. Will the company publish its protocol: which problems, how many attempts, what resources? Will the solutions be verified formally or by humans, and who exactly does the checking? Will it say anything about the benefit to working mathematicians, rather than only the percentage of problems closed? That is precisely what the letter puts first.

For engineers and buyers of AI infrastructure, the takeaway is practical. If you are building RL pipelines for science, the most valuable part of a math challenge is not the leaderboard but the ready-made combination of environment, verifier and scalable rollout. It can be carried over to chemistry, materials science and provable programming, where verification can also be automated. If you are choosing a vendor, ask not for a benchmark table but for a reproducible report with GPU-hour costs and a description of the verification. The Fields medalists' criticism is precisely that a bare table proves nothing.