Why Is AI Advancing So Rapidly in Mathematics?

🙋
I’m Luhui Dev, a developer who has been breaking down Agent engineering and exploring how AI can be applied in education.
I focus on Agent Harness, LLM application engineering, AI for Math, and the productization of education SaaS.

How many pairs of points can be exactly one unit apart if you place a set of points in the plane?

This question, which even a secondary school student can understand, has troubled mathematicians for nearly eighty years. The Hungarian mathematician Paul Erdős conjectured that, as the number of points increases, the maximum number of these “unit-distance pairs” grows only slightly faster than the number of points.

In May 2026, OpenAI announced a breakthrough: an internal reasoning model had found a new construction that disproved this long-standing conjecture. According to the published materials, external mathematicians checked the proof. It did not settle the final optimal answer to the unit distance problem, but it showed that a widely believed growth pattern was wrong. [1]

The method was also surprising: the model applied deep tools from algebraic number theory to a seemingly intuitive problem in plane geometry.

Meanwhile, Google DeepMind introduced the mathematics research agent Aletheia in 2026, demonstrating efforts to have AI repeatedly propose, verify, and revise proofs and participate in professional mathematical research. [2]

Why is AI, which once made mistakes in basic arithmetic, beginning to reach the frontiers of mathematical research? Five developments help explain the change.

1. AI can keep learning from failed mathematical attempts

When training a model to write, it is difficult to assign every article a single, objective score. Whether prose is elegant or an idea insightful often depends on the reader and the context.

Many mathematical tasks offer clearer feedback.

A calculation can be checked against an answer; a construction satisfying particular conditions can be tested by a program; a proof written in a formal language can be checked by a proof assistant.

This makes mathematics an important training ground for reinforcement learning. The model attempts a problem, the system rewards it based on the result, and the model is adjusted so that effective strategies become more likely. With extensive practice, behaviors such as breaking down problems, checking answers, and changing approaches may emerge.

The DeepSeek-R1 research demonstrated that reinforcement learning with verifiable outcomes can substantially improve a model’s mathematical reasoning. [3]

It is a little like giving a student a practice room that is always open: they can keep working on problems and promptly learn which attempts succeeded.

2. AI can devote more compute to solving a problem

Hard mathematical problems rarely lead to an answer along a straight path.

Anyone who has worked on geometry problems knows the experience: several auxiliary lines lead nowhere, then a different construction suddenly unlocks the proof. Mathematical research is much the same. Trying, getting stuck, backtracking, and recombining ideas are part of the work.

Some AI systems now devote more computational resources to these processes. They can generate multiple candidate approaches, split a problem into subproblems, check key steps, and continue exploring promising directions.

This is known as increasing “inference-time compute”: after training is complete, a model can still improve its chances of success by spending more computation on a particular problem.

Aletheia uses an iterative process of generating, verifying, and revising solutions, together with tools for handling long proofs and the literature involved in professional research. [2]

Think of navigating a maze: try several routes, record the dead ends, and concentrate on branches that still look promising.

3. Language models and verification tools collaborate in agent systems

Large language models are good at proposing candidates. A problem may bring to mind a theorem, a transformation, or an auxiliary construction worth trying.

That flexibility is valuable, but a model can also overlook assumptions, misuse a theorem, or treat an unproved claim as an established fact.

Symbolic reasoning systems and proof assistants can take on another part of the work.

For example, AlphaGeometry has a language model propose potentially useful geometric constructions, then uses a symbolic reasoning engine to derive consequences. This connects flexible candidate generation with rigorous rule-based computation. [4]

Proof assistants such as Lean require statements and deductions to be written precisely so that a computer can check whether every step follows the logical rules. AlphaProof used such a formal environment for learning and proof search. [5]

This is like giving exploration a strict inspector. The model can try bold ideas while the verification system guards the logical boundaries.

The inspector has a limited remit, too: it can check a proof of a formal statement, but someone still needs to ensure that the statement accurately represents the original problem. Natural-language understanding, formalization, and proof checking must work together.

4. AI can generate practice material to ease the shortage of quality data

Humanity has accumulated a vast body of mathematical knowledge, but high-quality material suitable for training AI is not unlimited. Complete, correct proofs at an appropriate level of difficulty are especially valuable.

Some mathematical domains allow programs to generate new problems and constructions and then check the results. This creates opportunities to expand training material.

AlphaGeometry used programmatically generated geometry training examples, reducing its reliance on human-written proofs. [4]

This can create a cycle: the system generates problems, the model tries to solve them, tools filter for correct results, and qualified material feeds into further training.

The cycle may expose models to more structures and methods. But quantity is only one factor. Large numbers of repetitive or overly easy problems may not help with unfamiliar ones, while insufficiently screened solutions can amplify errors.

5. AI can try connecting different mathematical fields

Mathematical knowledge is highly dispersed. A geometry problem may require number theory; a combinatorial problem may conceal an algebraic structure. Finding these connections often requires crossing disciplinary boundaries.

Learning from a broad range of mathematical material gives large language models a basis for making such connections. Search tools can also help them examine candidate approaches and relevant literature at scale.

The new development on the unit distance conjecture illustrates the value of cross-field methods: the published proof brought algebraic number theory into discrete geometry and produced a previously undiscovered construction. [1]

Another example comes from a systematic investigation of the Erdős Problems database. Researchers used Gemini to examine 700 problems labeled “Open.” The reported progress included apparently novel solutions as well as old answers recovered from existing literature. The authors also emphasized the difficulty of establishing novelty and avoiding unintentional rediscovery of earlier work. [6]

This reveals two contributions mathematical AI can make: exploring new methods and helping researchers reorganize existing knowledge. Both can move work forward, but they need to be distinguished clearly.

Closing thoughts

Across these five developments, AI’s advantages center on the scale of exploration and the efficiency of feedback. It can try many candidates in a relatively short time, use tools to check results, and learn from success and failure during training. For mathematical problems with scattered knowledge and many possible approaches, this can expand the territory researchers are able to explore.

AI’s strength is allowing more ideas to be tried; human mathematicians bring judgment about which attempts deserve further attention and what understanding can be drawn from the results. Even after a statement has been proved, many questions remain: what structure actually makes the proof work? Which assumptions can be removed? Can the method explain a different class of phenomena? Research intuition, an understanding of theoretical context, and the ability to discuss and challenge ideas with peers all inform these judgments.

These strengths do not mean creativity can be assigned neatly to either side. AI can already suggest surprising constructions, and human mathematicians are also skilled at computation, search, and verification. Their division of labor will change as tools improve. The more useful comparison concerns a specific research task: who can more effectively propose directions, find errors, explain results, and advance the next step?

Mathematics needs correct results and results that can be understood and generalized. As AI expands the scope of exploration, a development worth watching is whether humans and AI can turn these discoveries into clearer concepts and theories with greater explanatory power.

References

1. OpenAI: An OpenAI model has disproved a central conjecture in discrete geometry (2026).
2. Feng et al.: Towards Autonomous Mathematics Research (2026).
3. DeepSeek-AI: DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning (2025).
4. Google DeepMind: AlphaGeometry: An Olympiad-level AI system for geometry (2024).
5. Google DeepMind: AI achieves silver-medal standard solving International Mathematical Olympiad problems (2024).
6. Feng et al.: Semi-Autonomous Mathematics Discovery with Gemini: A Case Study on the Erdős Problems (2026).