Jiuzhang Yunji Wins Terence Tao's Math Challenge with Perfect Scores in All Four Tracks

On September 11, 2026, 28 Fields Medalists jointly signed and published "Severe Misalignment of Artificial Intelligence in Mathematics," pointing out that AI companies treating the solving of difficult mathematical problems as a benchmark for measuring their own capabilities is "harmful to mathematical science and the mathematical community." The logic of this statement is that problem-solving is only a means and a proxy indicator, while what mathematical research truly pursues is conceptual understanding and insight.

Half a month later, at the end of September 2026, the SAIR Foundation co-founded by Terence Tao announced the results of the second stage of the Math Distillation Challenge. On the surface, this is a competition about AI doing mathematics, and it seems to overlap with the target of that statement's criticism; but the competition rules themselves give another answer: if the implication holds, the participating system must submit a Lean 4 formal proof; if the implication does not hold, it must submit a counterexample, that is, an algebraic structure that satisfies the premises but violates the conclusion; all submissions are checked one by one by a deterministic Lean verifier, and no points are awarded if verification fails.

In other words, the judging standard was rewritten from "accuracy rate" to "verifiability." This is the pivot for understanding the whole matter: what Terence Tao has always opposed is the path of "pursuing only correct answers and not caring about understanding," not AI doing mathematics itself. The competition's requirement of formal verification is itself an expression of a position.

Among the 294 participating teams, Yanbiao Lab (yanbiao.ai) from Jiuzhang Cloud achieved a perfect 200 points in all four tracks, with an overall leaderboard score of 221 points, far ahead of the other participants and ranking as the sole first place. The four tracks are formed by crossing two dimensions: the Solo track handles one problem at a time with a fixed per-problem budget, focusing on depth in attacking individual problems; the Marathon track handles 100 problems at once with a shared global budget, focusing on resource allocation and overall strategy.

What better illustrates the difficulty distribution is the Research Leaderboard outside the main leaderboard. The competition evaluation set is divided by difficulty into four levels: Normal, Hard, Extra Hard, and Order 5, where Order 5 equations involve about 62,000 laws and cover a large number of open problems that previously had no settled conclusion. The Research Leaderboard separately scores the most difficult portion selected from these Order 5 equations. Among all participating teams, only two teams scored above zero: Jiuzhang Cloud led with 17 points, the second place had only 3 points, and all the rest scored 0.

It is worth noting that in multiple tracks, the participating team used the same set of agents and methods. During iteration, the team found that it did not need to design separate agents and algorithmic tools, but instead built a proof graph containing billions of lines of Lean code, then compressed it to within the 500 KB Python code required by the competition. This is an engineering constraint and also a methodological choice: it means the results did not come from customized tuning for different tracks.