Xiaohongshu dots Model Claims an IMO 2026 Perfect Score, but Verification Is the Real Test

Xiaohongshu says its dots-note 3.0 model earned 42 points from 42 on the 2026 International Mathematical Olympiad problems. The Xiaohongshu dots model IMO 2026 perfect score claim covers all six problems, matching the seven human contestants reportedly tied for first place.

The result would move AI mathematical reasoning beyond the milestone established by Google DeepMind and OpenAI in 2025. Their experimental systems reached gold-medal level with 35 points, equivalent to five complete solutions. A perfect paper removes the last visible gap on this particular test.

However, dots-note 3.0 did not enter the student competition as a contestant. Xiaohongshu evaluated an internal model on the contest problems and reported the outcome through its dots research team. That distinction matters because a score alone does not reveal the model’s access, compute, prompting, selection process, or grading procedure.

The central question is therefore larger than whether an AI generated six convincing proofs. It is whether the evaluation was controlled closely enough to support meaningful comparison with human contestants, Gemini Deep Think, OpenAI’s experimental system, and future public models.

What Xiaohongshu Says dots-note 3.0 Accomplished

The reported 42-point result matters because it combines a perfect score with an end-to-end natural-language approach.

The dots team says dots-note 3.0 received the original LaTeX versions of the six IMO 2026 problems. LaTeX is a document format commonly used to express equations and technical notation. The system reportedly generated complete solutions without first translating each problem into a formal theorem-proving language.

According to the team’s model account, dots-note 3.0 received full credit on every problem. That produced the maximum possible score of 42 points. Each IMO problem carries seven points, and contestants solve three problems on each of two examination days.

The 67th IMO took place in Shanghai, where the examinations were held on July 15 and July 16. The 2026 regulations identify Shanghai High School as the examination venue and allocate four and a half hours for each daily session.

Xiaohongshu describes dots-note 3.0 as the lightest model in its dots3 family. The company has not disclosed enough technical detail to compare “lightest” through parameter count, active parameters, inference cost, memory use, or wall-clock performance. It has also not released the model weights.

The team says the model relies on recursive self-critique. In this process, the system reviews intermediate reasoning, identifies possible errors, and revises its approach before producing a final answer. The method is closer to repeated internal review than a single uninterrupted response.

That mechanism is relevant to olympiad mathematics because a promising solution can fail through one unsupported step. A model must find an argument, test its internal consistency, repair gaps, and express the final proof clearly enough for expert grading.

The reported result also differs from answer-only mathematics benchmarks. IMO problems require proofs, so reaching the right numerical conclusion is insufficient. Every essential inference must be justified, and a hidden flaw can reduce a seemingly complete solution to partial credit.

Xiaohongshu says dots-note 3.0 completed this process end to end. However, the public account does not yet provide a complete evaluation package covering all prompts, outputs, timestamps, retries, compute limits, and grader identities.

That missing package does not make the result false. It does mean readers should treat 42 out of 42 as a company-reported benchmark result awaiting fuller independent review.

Why a Perfect Score Changes the AI Math Race

Gold-medal performance showed that models could solve most of an IMO paper, while a perfect score claims they can survive every problem on the same attempt.

In 2025, Google DeepMind’s advanced Gemini Deep Think system solved five of six problems and scored 35 points. IMO coordinators graded its natural-language proofs under the same marking criteria used for student solutions.

The company’s Deep Think evaluation marked a sharp improvement from its 2024 system. AlphaProof and AlphaGeometry 2 had scored 28 points, but they relied on specialist formal languages and required days of computation.

Gemini’s 2025 system instead read official problem descriptions and wrote proofs in natural language within the competition’s four-and-a-half-hour time limit. It used parallel reasoning to explore and combine several possible approaches.

OpenAI also reported a 35-point result in 2025 from an experimental general-purpose reasoning model. Its announcement drew scrutiny because it arrived before the IMO’s coordinated disclosure process had finished. Google’s submission received explicit certification from IMO graders.

The distinction between an official medal and medal-level performance became easy to lose. AI systems were not regular student contestants, even when experts evaluated their answers. Their scores indicated how the work compared with contestant submissions, not that a machine had entered a national team.

The same precision applies to the Xiaohongshu dots model IMO 2026 perfect score. The meaningful claim is that dots-note 3.0 reportedly produced proofs worth 42 points under the team’s evaluation conditions. It did not displace a human contestant from the standings or receive a student medal.

Still, solving six problems would represent more than a seven-point increase. The final problem is often among the paper’s hardest, and a gold-level model can reach its total by solving the more accessible five. A perfect score requires success across the entire problem set.

Seven human contestants reportedly earned perfect scores in 2026. That places the claimed model result at the top of the human score distribution rather than slightly above the gold threshold.

It also changes the competitive target for other laboratories. Matching last year’s 35 points would no longer define the leading result. AI developers would instead face pressure to report complete papers, stable repeat performance, and controlled evaluations that distinguish capability from extensive search.

This pressure extends beyond Google and OpenAI. Developers of open-weight reasoning models now have a specific claim to test. A smaller model that can reproduce perfect performance would challenge the assumption that the strongest mathematical reasoning requires the largest proprietary systems.

The important phrase is “reproduce perfect performance.” One successful run can establish possibility, but it cannot establish reliability. Scientific and commercial users care about the probability of getting a correct proof when the answer is not already known.

That is where the competitive race moves next. The headline score is 42. The harder benchmark is whether another evaluator can obtain it again.

How Recursive Self-Critique Targets the Hardest Part of Proof Writing

Dots-note 3.0 appears designed to improve reasoning through revision, but the outcome depends on how candidates are generated, judged, and selected.

A standard language model generates text by predicting a sequence of tokens. Longer reasoning can help it decompose a problem, yet length alone does not guarantee correctness. The model can repeat an early error with increasing confidence.

Self-critique adds a review stage. The model inspects a proposed argument, searches for gaps or counterexamples, and attempts a corrected version. Recursion means that this review can repeat across several levels rather than stopping after one critique.

This approach fits mathematical proof writing. A geometry construction can fail in a special case. An inequality can depend on an unstated sign condition. A combinatorics argument can count the same object twice. Rechecking creates opportunities to detect those failures before submission.

The method also resembles how experienced mathematicians work. They try several representations, reject attractive dead ends, test boundary cases, and rewrite the successful argument. The resemblance concerns workflow, not proof that a model understands mathematics like a person.

Xiaohongshu says its model handled the original LaTeX questions directly. If confirmed, this removes one source of human intervention. Earlier formal systems often required specialists to convert informal contest language into machine-readable statements before solving began.

Natural-language reasoning offers flexibility. It can engage with diagrams, informal descriptions, and proof strategies without requiring every concept to be encoded in a theorem prover. It also produces answers that human graders can read directly.

Formal methods offer a different advantage. A proof assistant can mechanically verify whether every logical step follows from accepted rules. Natural-language proofs remain vulnerable to persuasive wording that hides a genuine gap.

That creates the main technical opponent in this story: natural-language self-critique versus formally verified reasoning. The Xiaohongshu result supports the natural-language route, but it does not settle the contest.

A recursive critic can share the solver’s blind spots. If both stages come from the same model, the evaluator may approve an argument that reflects a common internal misconception. Multiple candidate solutions reduce this risk, but only when the selection method can distinguish subtle errors.

External verification can provide that distinction. A separate model, a formal proof system, or independent mathematicians can challenge the final answer. Each option changes the resources behind the reported score.

This is why evaluation details are part of the result rather than administrative fine print. Ten thousand sampled solutions followed by expert selection describe a different capability from one model producing one accepted answer per problem.

The same applies to prompt design. General instructions to check work differ from problem-specific hints written after researchers studied the questions. Both can generate valid proofs, but they measure different degrees of autonomous reasoning.

Compute also matters. Recursive self-critique can spend additional inference on difficult problems. That may be desirable for scientific work, where accuracy matters more than response speed. It still makes comparisons misleading unless laboratories publish their time and compute budgets.

A light model with a long reasoning budget can outperform a larger model restricted to one short response. Parameter count alone cannot explain the result. The entire inference system, including search, critique, verification, and selection, defines the capability.

This lesson extends to real research. A model helping an engineer check an argument should not merely generate a plausible answer. It should record assumptions, preserve rejected alternatives, and make its final evidence traceable.

Knowledge workers face a similar challenge when comparing model outputs with internal documents. A searchable engineering knowledge base can preserve the specifications and prior decisions needed to check a generated conclusion.

The dots approach is therefore interesting even before independent reproduction. It places revision at the center of difficult reasoning. What remains unproven is whether its critic is dependable when no olympiad grader can recognize the answer afterward.

The Xiaohongshu dots Model IMO 2026 Perfect Score Still Needs a Reproducible Evaluation

The score is impressive as a reported result, but the current evidence does not support treating it as an independently established model record.

The first verification question concerns problem access. A clean evaluation should establish when the model and researchers received the official questions, whether internet access was disabled, and whether any retrieved material entered the context.

IMO problems are new, which limits direct memorization. However, they draw on established methods, and online discussions can appear quickly after an examination. A timestamped run is therefore important.

The second question concerns human assistance. Researchers should disclose whether people selected prompts, supplied hints, rejected outputs, edited proofs, or chose among candidate solutions. A model can still deserve credit within a human-guided system, but that system should not be described as fully autonomous.

The third question is sampling. A single answer per problem tests direct reliability. Hundreds of attempts followed by selection test whether the model can produce a correct solution somewhere in its distribution. Those are useful but different measurements.

The fourth question is grading. IMO solutions often receive partial credit, and assigning seven points requires confidence that every necessary step is present. Independent graders should see the exact unedited outputs rather than reconstructed solutions.

The official IMO results archive is authoritative for human contestants and medal thresholds. It does not automatically validate a company’s separate AI evaluation. Any certification of dots-note 3.0 would need its own documented process.

This distinction has precedent. Google DeepMind stated in 2025 that IMO coordinators officially graded and certified Gemini’s solutions. The company also published the model’s proof documents. That disclosure did not reveal every implementation detail, but it created a stronger basis for examining the score.

OpenAI’s 2025 result illustrated the communication risk. Its early announcement created a dispute over timing and official status, even though expert graders reportedly considered the solutions gold-worthy. Coverage in Nature treated both companies’ work as a major reasoning milestone while preserving the differences between their evaluations.

Xiaohongshu can close much of its present verification gap by publishing a compact evidence package. It should include the six raw outputs, the full system instructions, run timestamps, compute limits, sampling counts, and grading records.

Releasing model weights would go further, but weights alone are insufficient. Researchers would also need the inference procedure because recursive self-critique may depend on orchestration outside the base model.

The team says dots-note 3.0 is expected to become open source. Until a license, repository, weights, and runnable evaluation code appear, that remains a future commitment rather than a feature readers can inspect.

Potential contamination also deserves careful treatment. The 2026 problems were unavailable during ordinary pretraining, assuming the model’s training cutoff preceded the event. Yet a post-event evaluation can still encounter solutions through retrieval, updated data, or human prompts.

No public evidence currently shows that this happened. The point is methodological: a perfect score demands stronger controls because there is no remaining scoring margin to expose hidden assistance.

There is also a benchmark saturation problem. Once several laboratories optimize directly for olympiad performance, the IMO becomes less useful as a general measure of intelligence. Teams can train specialized reasoning policies without improving reliability across other domains.

Research on IMO proof benchmarks attempts to address this problem with expert-vetted tasks and separate evaluation sets. Even those benchmarks cannot fully predict performance in scientific research, where questions are open-ended and correct answers may not yet exist.

A perfect olympiad score therefore supports a narrow but important conclusion. The model can reportedly generate highly sophisticated competition proofs under the team’s setup. It does not establish flawless mathematics, autonomous discovery, or dependable reasoning in every professional setting.

That restraint protects the result rather than diminishing it. Clear evaluation boundaries allow later replications to strengthen the claim. Inflated language makes every missing detail look like contrary evidence.

Who Faces Pressure if the Claim Holds Up

A reproducible result would pressure proprietary frontier labs, open-model developers, and benchmark designers at the same time.

Google DeepMind faces the most direct comparison because Gemini Deep Think established the best-known officially graded natural-language result from 2025. Its 35 points created the baseline that dots-note 3.0 claims to exceed by solving the remaining problem.

Google can respond through a 2026 evaluation under similarly documented conditions. A matching perfect score would show that the achievement is no longer isolated. A lower score would still be difficult to interpret without equivalent compute and prompting rules.

OpenAI faces similar pressure, particularly because its 2025 system was described as a general reasoning model rather than an olympiad-specific solver. A new comparison would test whether general-purpose scaling keeps pace with specialized self-critique systems.

Open-weight model developers face a different challenge. If Xiaohongshu releases dots-note 3.0, outside teams could inspect and adapt its reasoning process. That would make the result relevant to researchers who cannot access experimental systems from the largest US laboratories.

The model’s reported size class is especially important here. Xiaohongshu calls it the lightest member of dots3, but has not supplied the metrics needed to evaluate efficiency. Publishing parameter and compute data could turn a score story into an architectural one.

Benchmark designers also face pressure. A 42-point model leaves no room for future improvement on that single paper. They will need broader tests that measure consistency across unseen problems, sensitivity to perturbations, proof verification, and resource use.

One useful design would run models across multiple fresh competitions with fixed compute. Another would alter problem statements without changing their mathematical content. A system relying on surface patterns would become less stable under those changes.

Researchers should also compare pass rates across repeated runs. If a model solves a problem once in 100 attempts, it has discovered a valid path but remains unreliable. If it succeeds 90 times, the capability is much closer to operational use.

The human comparison requires equal care. Seven contestants reportedly achieved 42 points in Shanghai, but people and models bring different resources. Students have bounded exam time and no computers, while an AI system may use large data centers and extensive parallel sampling.

That does not make the model comparison meaningless. It means the score measures output quality under stated conditions, not equal cognition or equal resource use.

For developers, the broader implication concerns reasoning infrastructure. The winning component may not be a larger base model. It may be the loop that allocates time, generates alternatives, criticizes errors, and decides when a proof is ready.

For enterprise buyers, that suggests a better evaluation question. Instead of asking which model tops one leaderboard, ask whether the complete system can show its sources, repeat its result, detect failure, and operate within acceptable limits.

For individual users, the same principle applies to AI-assisted research. A model’s polished answer should be the beginning of verification, not its substitute. Preserving source material in a personal knowledge base makes it easier to test claims against the documents that support them.

The Xiaohongshu dots model IMO 2026 perfect score claim puts that principle under a bright spotlight. The better the output looks, the more important its provenance becomes.

Three Signals That Will Determine Whether the Result Lasts

The next stage is not another announcement but a sequence of disclosures and replications that can turn a company claim into durable evidence.

The first signal is publication of the complete six-problem record. Xiaohongshu should release the exact model outputs as generated, without editorial repair. Independent olympiad graders can then identify any hidden gaps and assign scores from the same text.

This release should include the prompts and instructions. A general instruction to solve and verify each problem would support the autonomy claim. Detailed hints, manual interventions, or selective resubmission would narrow it.

If independent graders also award 42 points, the central score claim becomes substantially stronger. If they find a major proof gap, the story shifts from perfect reasoning to evaluation inconsistency.

The second signal is a reproducible dots-note 3.0 release. The team has indicated that the model is expected to become open source, but users need more than downloadable weights.

A meaningful release should identify the license, parameter configuration, tokenizer, inference settings, critique loop, sampling policy, and compute requirements. It should also provide scripts that recreate the evaluation without undisclosed services.

Outside replication would test whether the published system generates comparable proofs rather than merely resembling the internal model. Stable results across several laboratories would strengthen the claim that recursive self-critique transfers beyond one controlled run.

Failure to release would not disprove the score. It would leave the efficiency and openness portions of the narrative unsupported.

The third signal is performance on new, independently controlled mathematics. IMO 2026 is already public, so every later run carries growing contamination risk. The strongest follow-up would use unseen problems delivered directly to evaluators.

Those tests should cover more than olympiad patterns. They should include proof checking, counterexample construction, ambiguous assumptions, and tasks where the proposed theorem is false. A system that always tries to prove the prompt can fail badly on adversarial research questions.

Repeated trials should accompany headline scores. Pass rates, compute budgets, and error categories would reveal whether dots-note 3.0 is consistently capable or occasionally successful after extensive search.

The same tests should compare natural-language and formal verification routes. Self-critique may generate creative candidate proofs quickly, while a theorem prover checks the final logic. A hybrid system could outperform either route alone.

A successful release and replication would reinforce the article’s central judgment: Xiaohongshu has moved into the leading group of mathematical reasoning developers. Weak disclosure or unstable reproduction would reduce the result to an intriguing internal demonstration.

Either outcome will teach the field something useful. A verified 42-point paper would establish a new standard for natural-language mathematical reasoning. A disputed paper would expose how little a benchmark number says without a complete evaluation trail.

The Xiaohongshu dots model IMO 2026 perfect score should therefore be read as both a capability claim and a verification challenge. Developers can watch for raw proofs, runnable code, and blind evaluations. Researchers can ask whether the system succeeds repeatedly when neither the model nor its operators have seen the problems.

For everyone using AI to reason over technical material, the practical lesson is immediate. Save the source, inspect the assumptions, and demand a reproducible path from question to conclusion. If Xiaohongshu supplies that path, its perfect score will matter far beyond one olympiad paper.