Google DeepMind’s AI Solved Dozens of Open Math Problems on Its Own

Google DeepMind is making progress in one of artificial intelligence’s most demanding areas: original mathematical research. Its AI systems have begun solving open problems that were previously unanswered, including questions drawn from established collections of mathematical conjectures. Unlike systems designed mainly to answer familiar questions, these research agents can generate possible proofs, test their reasoning and revise solutions with limited or no human intervention.

The developments are significant because mathematical research requires more than producing an answer that looks plausible. A result must follow logically from its assumptions, and even a small error can invalidate an entire proof. DeepMind is addressing this challenge through systems that combine advanced language models with automated verification tools. Its work includes Aletheia, powered by Gemini Deep Think, and AlphaProof Nexus, which uses formal proof search to investigate open mathematical questions.

Key Takeaways

• Google DeepMind has developed AI systems capable of making progress on unsolved research-level mathematics problems.
• AlphaProof Nexus reportedly solved nine of 353 attempted Erdős problems and proved 44 of 492 open conjectures from the Online Encyclopedia of Integer Sequences.
• Formal verification tools such as Lean help check whether a mathematical proof is logically valid.
• Aletheia, powered by Gemini Deep Think, has also demonstrated autonomous research results and contributions to human-AI research.
• The breakthroughs are meaningful, but they do not establish that AI has solved the most famous landmark problems in mathematics.

What Did Google DeepMind’s AI Solve?

One of the clearest examples is AlphaProof Nexus, a framework designed to help AI systems search for formal mathematical proofs. In a research paper published in May 2026, DeepMind researchers described an agent that autonomously solved nine of 353 attempted problems associated with mathematician Paul Erdős. Some of the questions had remained open for decades.

The same research reported that the system proved 44 of 492 open conjectures drawn from the Online Encyclopedia of Integer Sequences, a major mathematical resource containing sequences and related conjectures. The successful results were published alongside formal proofs and supporting material, allowing other researchers to inspect and verify the work.

These numbers need context. The system did not solve every problem it attempted: most of the Erdős problems in the test remained unsolved. Its achievement was finding valid proofs for a subset of difficult questions without researchers supplying the solutions in advance.

The results nevertheless suggest that AI can do more than reproduce established mathematical knowledge. Under the right conditions, it can search through possible arguments and identify solutions that add to the existing body of mathematical results.

How AlphaProof Nexus Works

AlphaProof Nexus uses AI to propose mathematical reasoning and a formal proof system to check whether the proposed steps are valid. The research describes agents that generate candidate proofs, receive feedback from the Lean proof assistant and continue searching when a candidate fails.

Lean is a programming language and theorem prover used to express mathematical statements in a precise form. Its verification process checks whether a proof follows from accepted definitions and logical rules. This provides a stronger safeguard than asking another language model whether an answer appears correct.

The process can be understood in four stages:

1. Generate: The AI proposes a possible proof or a sequence of mathematical steps.
2. Check: The candidate is tested using a formal verification system.
3. Revise: If the proof fails, the agent changes its approach and tries again.
4. Verify: A successful proof is checked mechanically against the formal statement.

This feedback loop is important because AI models can produce convincing but incorrect explanations. Formal verification gives researchers a way to distinguish a valid proof from a fluent-sounding mistake.

The system also uses multiple agents and search strategies to explore possible solutions. Rather than relying on a single response, it can investigate different approaches and use feedback to focus its efforts.

Aletheia Takes AI Into Research-Level Mathematics

AlphaProof Nexus is not DeepMind’s only effort in this area. The company has also developed Aletheia, a mathematics research agent powered by Gemini Deep Think.

In a February 2026 research update, DeepMind described Aletheia’s work on professional research questions across mathematics, physics and computer science. The agent generates candidate solutions, checks them using a natural-language verifier, and revises or abandons approaches that appear flawed. It can also search existing literature to check references and reduce the risk of presenting known results as new discoveries.

DeepMind reported several different types of contribution. One paper described research produced autonomously by AI on structure constants called eigenweights in arithmetic geometry. Other projects involved human-AI collaboration, including work on mathematical bounds and generalizations of existing results.

Aletheia also participated in an evaluation covering 700 open questions from Bloom’s database of Erdős conjectures. DeepMind reported autonomous solutions to four open questions in that evaluation, with work on one problem contributing to a generalization reported in a research paper.

These results illustrate why the phrase AI solved a problem needs careful interpretation. Some work is produced autonomously, some is developed with human researchers, and some AI contributions are intermediate steps in a larger research project. Each category can be useful, but they represent different levels of independence.

Why Solving Open Problems Matters

Mathematical research often progresses slowly because a difficult question may require researchers to explore many possible approaches before discovering a productive one. A useful AI system could reduce that search time by generating candidate arguments, checking technical details and exploring directions that people have not yet considered.

The potential benefits extend beyond pure mathematics. Mathematical reasoning underpins computer science, cryptography, physics, engineering, statistics and parts of modern machine learning. Better tools for finding and checking proofs could help researchers investigate questions in these fields more efficiently.

AI could also make specialised mathematical knowledge easier to work with. A researcher may use an agent to examine a conjecture, search related literature, test intermediate steps and identify a proof strategy worth developing further. Instead of replacing the researcher, such a system could handle parts of the investigation that consume time without requiring the same level of creative judgement.

However, finding a valid proof is not the only measure of mathematical importance. A result may be correct but modest, while another may introduce a powerful technique or transform an entire field. Establishing the significance of a discovery requires understanding how it relates to previous work and what new possibilities it creates.

What About Famous Unsolved Problems?

DeepMind’s results should not be confused with a solution to the most famous open problems in mathematics, such as the Riemann hypothesis. The company’s February 2026 account explicitly said that it was not claiming a major advance or landmark breakthrough under its proposed classification of AI-assisted mathematical results.

This distinction matters because headlines about AI solving difficult mathematics can make different achievements sound equivalent. Solving a long-standing but narrowly defined conjecture is a genuine accomplishment. Resolving a landmark problem with far-reaching consequences would be a different level of achievement.

Independent scrutiny remains essential. Researchers need to establish that a formal statement accurately represents the original mathematical question, that the proof is valid and that the result is genuinely new. A computer can verify a proof relative to a formal statement, but experts may still need to confirm that the statement captures the intended problem and that the discovery has been described appropriately.

The most responsible interpretation is that AI systems are beginning to make verifiable contributions to research mathematics, rather than that machines have replaced mathematicians or solved mathematics as a whole.

How This Changes the Role of Mathematicians

AI research agents could change the day-to-day work of mathematicians by taking on tasks such as testing conjectures, exploring possible proofs and checking lengthy arguments. This may allow researchers to spend more time choosing important questions, interpreting results and deciding which discoveries are worth pursuing.

The change also raises questions about attribution and transparency. If an AI system generates a proof independently, researchers need to explain what the system did, which tools it used and how the result was verified. If a human supplies the core idea and AI completes the proof, the contribution should be described differently.

DeepMind has proposed a classification for AI-assisted mathematical research based on the significance of the result and the degree of AI involvement. Such distinctions could help the mathematical community assess claims more consistently as AI-generated research becomes more common.

Access to results is another important issue. Publishing proof files and formal verification materials allows other mathematicians to inspect the work rather than relying entirely on a company’s announcement. Open evaluation and reproducibility will be increasingly important if AI systems are to become trusted research tools.

The Bigger Picture

Google DeepMind’s work represents a shift from using AI primarily as an assistant for solving known problems toward using it as a tool for discovering new mathematical results. AlphaProof Nexus demonstrates how AI-driven search can find formal proofs for a subset of open problems, while Aletheia shows how a more general research agent can contribute to mathematical investigations through autonomous and collaborative workflows.

The deeper breakthrough is not simply that an AI model can produce advanced equations. It is that systems can combine generation, search, revision and verification to produce results that researchers can examine. Formal proof checking provides a valuable safeguard against errors, although it does not by itself establish that a discovery is important or entirely new.

Looking Ahead

The next stage will depend on whether these systems can solve more difficult problems consistently, produce clearer explanations and demonstrate value across a wider range of mathematical fields. Researchers will also need transparent reporting of failed attempts, computational costs, human involvement and independent verification. These details will help determine how much of the work can be reproduced and how broadly the methods can be applied.

For now, AI should be viewed as an increasingly capable participant in mathematical research rather than a replacement for human expertise. DeepMind’s results show that machines can independently find some valid solutions to questions that remained open, but selecting important problems, interpreting discoveries and establishing their broader significance will continue to require careful mathematical judgement.

Frequently Asked Questions

1. What is AlphaProof Nexus?

AlphaProof Nexus is a Google DeepMind research framework that uses AI agents and formal proof search to investigate mathematical problems and generate proofs that can be checked by a computer.

2. How many open Erdős problems did it solve?

DeepMind’s May 2026 research reported that its full-featured agent solved nine of the 353 Erdős problems it attempted.

3. What is Lean, and why is it important?

Lean is a formal theorem-proving system that checks whether a proof follows from specified definitions and logical rules. It helps distinguish valid formal proofs from explanations that merely appear convincing.

4. Did Google DeepMind’s AI solve the Riemann hypothesis?

No. DeepMind did not claim to have solved the Riemann hypothesis or another landmark problem of comparable significance in its reported results.

5. Will AI replace mathematicians?

There is no evidence that these systems can replace mathematicians. They can help explore ideas, search for proofs and check arguments, while human researchers remain important for evaluating significance, interpreting discoveries and directing research.

6. Why is this breakthrough important?

It shows that AI can contribute to original research by finding verifiable solutions to some previously open mathematical questions, potentially speeding up work across mathematics and related scientific fields.