AI and IMO Math Problems: What the Results Show

Yes—AI systems can solve many International Mathematical Olympiad (IMO) problems, and recent results show that leading systems can reach medal-level performance on some official contests. That does not mean AI solves every olympiad problem reliably: performance depends on the model, the year, the tools allowed, and how solutions are checked.

What the IMO measures

The IMO is an annual international mathematics competition for secondary-school students. Each participating country sends a small team, and contestants work individually on six proof problems over two days. The problems draw on areas such as algebra, geometry, number theory, and combinatorics; they are designed to reward mathematical reasoning rather than advanced university coursework.

Each problem is worth seven points, for a maximum of 42. Human solutions are judged against rigorous proof standards by coordinators and a jury. A correct numerical answer alone is not enough: a solution must establish why the result follows. Medal thresholds vary by year, because they depend on the distribution of scores. This matters when comparing an AI score with a medal boundary.

What the recent AI results mean

In 2024, Google DeepMind reported that its AlphaProof system and AlphaGeometry 2 together solved four of the six problems from that year’s IMO, earning 28 points and a silver-medal-level result under the contest’s scoring. The systems were not one general-purpose chatbot: AlphaProof used formal mathematics and reinforcement learning, while AlphaGeometry 2 specialized in geometry. Their results were assessed by IMO coordinators.

In 2025, Google DeepMind reported that Gemini with Deep Think solved five of six problems, scoring 35 points and reaching the gold-medal threshold. The solutions were produced under contest-like time limits and assessed by IMO graders. This was a major improvement, but it should be described precisely: it was a reported result by a particular system on a particular contest, not evidence that every AI can reproduce it or that every problem has become easy for machines.

These results show that AI can now produce complete, persuasive solutions to exceptionally difficult problems. They do not establish that AI has human-like understanding, that its reasoning is always dependable, or that it can solve any problem on demand. A contest score is a useful benchmark, not a universal guarantee.

How AI solves proof problems

There are several broad approaches. A language model may generate a proof in ordinary mathematical language, drawing on patterns learned from examples. Other systems search through possible deductions or use a formal proof assistant, where each step must follow from explicitly stated rules. Hybrid systems combine these methods: a model proposes ideas, computational tools explore them, and a formal checker verifies a proof.

Formal verification is important because fluent prose can conceal a gap. A proof assistant can check that each claim follows from earlier statements, but translating a creative olympiad argument into its formal language is itself difficult. Systems may also use specialized components for particular domains, such as geometry, rather than relying on one uniform method for all six problems.

A typical successful process can involve:

• Interpreting the wording and identifying what must be proved.
• Trying transformations, lemmas, constructions, or known mathematical techniques.
• Searching for a proof path, sometimes with computational guidance.
• Checking the final argument for logical gaps and matching it to the exact question.

That outline is not a claim that every system follows the same sequence. The architecture and permitted tools differ, and public reports may not expose every technical detail.

Why six problems remain difficult

IMO problems are deliberately varied. A model that is strong at algebraic manipulation may still struggle to invent the right geometric construction or discover a subtle invariant in a combinatorics problem. Even within a subject, a small change in wording can invalidate a familiar template.

Proof writing also has a different standard from getting a multiple-choice question right. A model can make a promising conjecture, test examples, or produce an outline without proving the general statement. It may overlook an exceptional case, assume what it needs to prove, or introduce an unjustified step. On the IMO, such a gap can cost points even when the final claim is true.

Finally, scores aggregate results across six problems. A high total can coexist with one unsolved problem, as the 2025 report illustrates. It is more informative to ask which problems were solved, how the proof was verified, and under what conditions than to treat a medal label as an all-purpose measure of mathematical ability.

AI versus human contestants

Medal comparisons provide an understandable benchmark, but AI and students do not necessarily compete under identical circumstances. Human contestants work with handwritten solutions and a set contest schedule. AI evaluations may involve different interfaces, system preparation, computing resources, or tools. Reports can impose time limits and use human graders, but comparisons still need to state the exact setup.

Human contestants also bring skills that a score does not capture: explaining an idea to another person, choosing what to study, responding to hints, and developing mathematical judgment over time. Conversely, computers can search or check many candidate steps quickly. The comparison is most useful as a measure of performance on a defined set of problems, not as a complete ranking of minds.

There is a further distinction between solving and explaining. An AI may produce a valid proof that is hard for a student to follow, or a clear-looking explanation with an error. Readers should separate readability from correctness and verify important arguments independently.

What the results are useful for

Strong olympiad performance suggests that AI can assist with tasks such as generating proof ideas, checking algebra, finding counterexamples, or explaining a known solution. Used carefully, it can be a study aid: ask for a hint before requesting a full proof, compare its argument with an official solution, and test each inference rather than copying the answer.

For teachers and researchers, the results are evidence that mathematical AI is improving on demanding reasoning benchmarks. They also expose weaknesses: a system may solve a celebrated problem yet fail on a nearby variation, or generate a convincing but invalid derivation. Independent checking remains essential, especially when an answer will be used in research, assessment, or instruction.

For a student, a practical way to use an AI explanation is to ask it to identify the key lemma, justify a non-obvious step, and show where each condition in the problem is used. Then try to reproduce the proof without looking. That tests understanding better than treating the output as an answer key.

Frequently Asked Questions

Has AI won an IMO gold medal?
In 2025, Google DeepMind reported that Gemini with Deep Think scored 35 points on the IMO and reached the gold-medal threshold. This describes that system’s reported performance, not a medal awarded to a human contestant or a guarantee for other AI models.

Did AI solve every problem on the IMO?
No. The reported 2024 result was four of six problems, and the reported 2025 result was five of six. A strong overall score does not mean all problems were solved.

Can ChatGPT solve IMO problems?
Some general-purpose chatbots can solve or help with certain olympiad problems, but performance varies by model, problem, and available tools. A polished response is not proof that the reasoning is correct; check the argument line by line.

How are AI IMO answers judged?
Contest solutions are evaluated for mathematical correctness and completeness, not merely for the final answer. Recent headline results were assessed by IMO coordinators or graders, as described in the reporting for those competitions.

Does an AI medal-level score mean AI understands mathematics?
It demonstrates impressive performance on a specific set of formal problems. Whether that amounts to understanding in the human sense is a broader question, and the score alone cannot settle it.

Related

• GPT Meaning: Generative Pre-trained Transformer Explained
• Best Cloud Storage for Large Photo Libraries
• Compare Nuclear, Coal, Gas, Wind and Solar Power