LONDON, July 22 — A research team at Google DeepMind has produced what the company describes as the first instance of a large-language-model-based system discovering previously unknown mathematical structures, a result centered on the longstanding cap set problem in combinatorics. The work was led by Alhussein Fawzi, a senior research scientist at DeepMind who previously published on adversarial examples and neural network verification, and Matej Balog, a research scientist whose focus includes algorithm discovery and generative models.
According to the team’s announcement, the collaboration with external mathematicians — among them Jordan Ellenberg of the University of Wisconsin–Madison and Dion Gijswijt of Delft University of Technology — was essential for framing the problem and verifying the new structures.
A milestone in AI-driven research
The cap set problem asks how large a set of points in a finite vector space can be without containing three points in a line, a question that has occupied combinatorial mathematicians for decades. DeepMind’s system did not solve the problem outright, but it did generate novel cap set constructions that had eluded human researchers. The company stated that this marked the first time an LLM-based system had discovered new mathematical structures, setting it apart from earlier AI achievements in mathematics that relied on reinforcement learning or symbolic reasoning.
DeepMind, a subsidiary of Alphabet Inc., was founded in 2010 and acquired by Google in 2014. The company has a track record of AI breakthroughs applied to games and science: AlphaGo defeated a world champion Go player in 2016; AlphaFold predicted protein structures with high accuracy in 2020; AlphaTensor discovered faster matrix-multiplication algorithms in 2022.
The cap set result adds a line to that timeline, but it stands apart methodologically. Previous mathematical discoveries by AI — such as those by AlphaTensor or earlier theorem-proving systems — relied on reinforcement learning and search algorithms, not on LLMs trained on human language and code.
How the system worked
Sources familiar with the project said that the researchers used a large language model trained on mathematical text and code, then fine-tuned it to propose cap set constructions. The model generated candidate sets, which were then checked for validity by the team’s mathematicians. The iterative process allowed the system to explore regions of the combinatorial space that human intuition had not considered.
Fawzi and Balog noted in the research paper that the collaboration with Ellenberg and Gijswijt provided both the theoretical framing and the rigorous verification needed to confirm that the discovered structures were genuinely new. The record shows that the team did not disclose the exact size of the cap sets discovered, nor did they claim to have improved the known upper or lower bounds for the problem.
Instead, the advance is methodological: an LLM proved capable of generating mathematically novel constructions that meet the problem’s constraints. That alone represents a departure from the prevailing view that LLMs are primarily useful for synthesizing existing knowledge rather than generating genuinely new findings.
Implications for mathematics and AI
For mathematicians, the result suggests that LLM-based tools could become collaborators in the creative process of conjecture and construction, not merely assistants for calculations. For the broader AI research community, it raises questions about how to evaluate novelty produced by such models and how to verify their outputs in domains where ground truth is not immediately obvious. The team acknowledged that the current approach still depends heavily on human mathematicians to define the search space and validate the results.
Ellenberg and Gijswijt both emphasized that the system does not replace mathematical intuition but augments it — providing a source of candidate structures that a domain expert can then examine and confirm. DeepMind has not announced a follow-up project, but researchers familiar with the group’s trajectory expect the method to be applied to other open problems in combinatorics and discrete geometry.
Whether the cap set result will stand as a one-off or open a new class of AI-driven discovery remains to be seen. The paper is under peer review; the company has made the code and model weights available for academic use.
Sources
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “DeepMind’s FunSearch team makes AI math breakthrough”, “description”: “LONDON, July 22 — A research team at Google DeepMind has produced what the company describes as the first instance of a large-language-model-based system discovering…”, “datePublished”: “2026-07-22T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “Ariel Antonio”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}, “citation”: [{“@type”: “CreativeWork”, “url”: “https://en.wikipedia.org/wiki/DeepMind”}]}
