Xiaohongshu achieves significant AI milestone

A new milestone in advanced artificial intelligence reasoning has been achieved as Chinese technology company Xiaohongshu, or RedNote, recently announced that its dots-note 3.0 model scored a perfect 42 points on this year's 67th International Mathematical Olympiad, making it the first AI system in IMO history to achieve both a flawless score and a gold-medal-level result.

The result makes dots-note 3.0 only the second AI model after Google DeepMind's Gemini Deep Think to receive an officially graded gold-level result. Gemini scored 35 points in 2025 after solving five of the six problems, matching that year's gold-medal cutoff.

Unlike many mathematical benchmarks that only ask for a final answer, the IMO requires complete proofs. A solution can lose points if any step is unsupported, making the contest a test of abstract reasoning as well as the consistency of a model's argument.

"Such competitions offer a useful test because the problems are newly written and unavailable online before the event," said Jasper Yao, a research scientist at a major US-based AI laboratory.

As online material can quickly enter model-training data, unseen questions help reveal whether a model can reason beyond material it may have encountered, Yao said, adding: "The value of taking part in a math contest is that it tests how well a model can perform when it has no prior knowledge of the problems and has never seen them before. That is also why major model developers compare their models' mathematical capabilities."

This year, the six IMO problems in question covered algebra, geometry, number theory and combinatorics, with each worth seven points. For the AI evaluation, human intervention was prohibited, including adjusting the model's reasoning, adding hints, correcting answers or selecting among outputs, according to the organizer.

Xiaohongshu said its dots-note 3.0 received the organizers' original LaTeX-formatted problems and solved them through an agent-based system combining natural-language reasoning with Python code execution.

"With no human translation or rewriting of the questions required, the model can completely critique its own work, identify errors or weak steps in its reasoning, and refine its initial proofs before producing the final answers," the company said.

The submissions were then reviewed by IMO graders, who awarded the model a seven-point full score for each problem.

A perfect score was rare even among human contestants. Seven of the more than 600 students listed in the official results received 42 points, or about 1 percent. The gold-medal cutoff this year was 29 points, placing the model's reported results 13 points above the threshold.

"We were surprised to find that the model produced a correct and elegant proof," said Liu Hanzuo, a two-time gold medalist at the Chinese Mathematical Olympiad. "Its structure was compact and its logic natural. Even among IMO solutions, it would be considered concise and elegant."

Wang Qiantong, another CMO gold medalist, also highlighted the model's solution to Problem 3 of this year's IMO.

While contestants commonly turned the problem into one about connectivity in graph theory, dots-note 3.0 used an induction argument that directly addressed the hardest part of the question, Wang said, adding, "Its approach is ingenious and concise, and that took me by surprise."

Xiaohongshu said the model used in the evaluation was an internal version of dots-note 3.0 and that it expects to open-source the model to the AI community in the near future.

lijiaying@chinadaily.com.cn