First, let's clarify what these 10 problems are.
They are not math competition problems or classic puzzles. They are unpublished research-level math problems from the FirstProof challenge, whose participants include professional mathematicians and math PhDs. The requirements: the answer must be correct, and it must be accompanied by a rigorous mathematical proof that can withstand peer review.
Aletheia, DeepMind's latest math AI, solved 6 of them.
Five were rated by external experts as "ready for publication with minor revisions." What does that mean? It means the quality of these solutions has reached the standard of real academic papers.
How Aletheia Works
Aletheia's underlying model is Gemini 3 Deep Think, which improves performance through extended test-time computation. But it's not just running a single large model; it's a multi-agent framework:
- Generator: Generates initial solution ideas
- Verifier: Checks proof steps for logical gaps
- Reviser: Modifies the proof based on verification feedback
- Integrates Google Search to access the latest mathematical literature
The idea behind this architecture is that verifying a proof is easier than generating one. In mathematical proofs, each step of reasoning is either correct or incorrect—this is a relatively objective judgment, much easier than generating content. So having the Verifier act as a gatekeeper can, to some extent, reduce the occurrence of "seemingly correct but actually flawed" solutions.
One other thing Aletheia does is noteworthy: when it encounters a problem it truly cannot solve, it explicitly outputs "no solution found" or times out, rather than forcing a wrong answer. The design of "knowing what you don't know" is far more important in mathematics than forced hallucination.
Comparison with OpenAI
In the FirstProof challenge, OpenAI also had an internal model participating.
Initial reports said OpenAI's model also solved 6 problems. After a review, Problem 2 was found to have a logical flaw, reducing the count to 5.
Aletheia under the same conditions: 6 solved, with 5 rated as publishable by external expert review.
| System | Problems Solved | Expert Review Result |
|---|---|---|
| Aletheia (DeepMind) | 6 | 5 "publishable" |
| OpenAI Internal Model | 5 (1 downgraded after review) | Not disclosed in detail |
The gap is small, but the direction is clear: math AI is entering a range where it can converse with top researchers.
91.9% on IMO-ProofBench
Beyond FirstProof, Aletheia achieved a 91.9% accuracy rate on IMO-ProofBench, a proof benchmark based on International Mathematical Olympiad problems.
For reference: the pass rate for typical university math majors on IMO problems is generally below 20%; for top math PhD students, it's 40-60%; only IMO silver medalists and above can consistently exceed 70%.
91.9% already surpasses the vast majority of human competitors who can participate in the IMO.
To be honest, the research team themselves didn't overhype it. There's a sentence in the paper worth quoting directly:
"Even with the Verifier mechanism, Aletheia's error rate is still higher than that of human experts."
This is an uncommon honesty: our benchmark scores are great, but we are still far from "replacing mathematicians."
Where the Line is Moving
A few years ago, large language models' performance in mathematics was: getting simple problems right, getting hard problems wrong, and failing at proofs.
Now the situation is: they can solve 60% of research-level problems, and most of those proofs are of publishable quality.
This isn't "better than humans," but rather "beginning to enter the range of discussing problems with human experts." The math community's reaction is more nuanced than the AI community's—some see it as a tool, others are starting to seriously think about what it means.
The interesting thing about mathematics is that its standard of correctness is objective. You either proved it or you didn't. The quality of code can be debated, but logical gaps in mathematical proofs can be detected by machines.
So mathematics is likely one of the first fields where the question "does AI truly understand or is it just statistical correlation?" can be clearly answered. Aletheia's 6/10 this time is a milestone in that journey of answering. Not the end.
Source: DeepMind's Aletheia Solves 6 Out of 10 Unpublished Research-Level Math Problems (InfoQ, CocoLoop, April 19, 2026)