Xiaohongshu's AI model dots-note-3.0 became the first AI to achieve a perfect score of 42 points at the 67th International Mathematical Olympiad (IMO 2026), with all six problems scoring 7 points. The model used natural language reasoning and Python-assisted analysis, employing a Proof-Verify-Refine self-correction loop, marking a significant advance in AI mathematical reasoning.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Xiaohongshu's dots-note-3.0 achieves the first official perfect score for an AI at IMO 2026, but the real breakthrough lies not in the score but in its recursive self-critique mechanism—potentially a key step toward general-purpose agents.
At IMO 2026, dots-note-3.0 submitted solutions that scored 7 points on all six problems, becoming the first AI model to achieve a perfect score in an official IMO evaluation. Previously, Google's Gemini Deep Think solved 5 out of 6 problems at IMO 2025, earning 35 points and meeting the gold standard. This year, seven human contestants worldwide achieved perfect scores, including China's Deng Leyan, Liu Che, and Zhang Bolun.
The value of a perfect 42 points lies in the fact that six complete proofs withstand step-by-step scrutiny. Proof problems differ from computational ones; a conclusion may be correct, but reasoning can have gaps. The IMO requires natural language proofs, and professional evaluation can expose hidden issues in long-chain reasoning.
dots-note-3.0's problem-solving process includes three stages: Proof, Verify, and Refine. Proof generates candidate proofs, Verify checks for gaps and provides suggestions, and Refine modifies and integrates solutions based on feedback. After completing one round, it can re-enter the checking process for multiple corrections.
This mechanism does not rely on formal systems like Lean. It directly reads LaTeX problems, uses natural language reasoning combined with Python for auxiliary analysis, and solves problems end-to-end in an agentic manner. The team calls this 'recursive self-critique' and views it as a prerequisite for recursive self-improvement (RSI).
In Problem 3 (combinatorial game theory and extremal proof), human contestants commonly convert it to a graph connectivity problem, but dots-note-3.0 chose induction: designing wooden segments that decrease by multiples, and proving by induction that no matter how Xiang Yu cuts, the total length Liu Bang loses is bounded by the same upper limit.
This shows the model has a certain autonomous exploration capability, able to bypass common approaches and find a more concise proof path. But a clever approach does not guarantee a perfect score; judges check the entire proof, and missing boundary conditions or key derivations can still result in point deductions.
IMO is a short-horizon, verifiable task with clear evaluation rules. Life tasks are longer and more ambiguous, such as renovation, travel, job hunting, etc., with many steps, changing requirements, and no unified scoring standards.
The team built VibeAgentBench, covering 10 domains including travel, finance, litigation, and renovation, with a total of 200 tasks, integrating 22 simulated service backends and 288 tool interfaces. Tasks last a median of 29 days, with the longest around 111 days, and are evaluated through 7,453 scripted events and over 12,000 check criteria.
These tasks, like IMO, test the model's ability to propose solutions, then check and revise based on feedback. In life scenarios, the model must examine more broadly: whether it understands the user, which preferences are worth preserving, and which needs have changed.
In the past, evaluating large models focused on the quality of the first response. Now, as tasks become longer and more complex, whether a model can pause during execution, check for deviations, and correct them in time becomes the new capability boundary.
dots-note-3.0's IMO perfect score is seen as a milestone validation, proving that the model can complete a closed loop of checking and correction in tasks with clear boundaries and verifiable results. The next exam has no fixed date and no standard answers.
This article's information primarily comes from a report by Jiqizhixin, where IMO results and model mechanisms are as stated by the reporting party and have not been independently verified by official sources. Some details, such as specific proof methods, come from third-party statements (e.g., CMO gold medalist Wang Qiantong) and are attributed.
dots-note-3.0's IMO perfect score is not just a technical breakthrough; it marks a paradigm shift in AI from single-shot generation to self-correction and recursive improvement, paving the way for long-horizon complex tasks.
Primary report
Primary source