AI Systems Pass Research-Level Math Benchmark, Ushering in Era of Verified 'Co-Mathematicians'

For years, a central question has haunted the intersection of computer science and higher mathematics: Can artificial intelligence actually reason, or is it simply executing highly sophisticated pattern-matching against problems it has already seen on the internet? To find out, a group of 30 leading mathematicians gathered at Harvard University's Center of Mathematical Sciences and Applications in early June 2026. Their goal was to administer a test that no AI could possibly have studied for, aiming to definitively map the boundary between human intuition and machine capability.[1][2]

The project, dubbed "First Proof," was designed with a brilliantly simple safeguard against data contamination. The organizers sourced ten original, unpublished mathematical problems directly from the active research of top-tier mathematicians. Because these specific lemmas and theorems had never appeared in textbooks, published papers, or on preprint servers like arXiv, they were entirely absent from the training data of any large language model. For an AI to solve them, it would have to demonstrate genuine, on-the-fly mathematical reasoning rather than regurgitating memorized proofs.[1][3]

The results, announced in mid-June, marked a watershed moment for the field. Across the four leading AI systems tested—which included frontier models from Google, OpenAI, and a specialized system from ETH Zurich—the machines successfully produced passing grades on seven of the ten research-level problems. The blind-grading panel of human experts noted that while some AI solutions required minor revisions, others were deemed "flawless," and in one instance, an AI model utilized a completely novel strategy that deeply impressed the referees.[1][3]

However, the benchmark also highlighted the enduring supremacy of human experts in specific domains. While the top-performing individual AI system solved six of the ten problems, the human mathematicians who contributed the questions were able to solve all of them. The consensus among the organizers was not one of impending doom for human mathematicians, but rather a recognition of "genuine capability." The AI systems proved they are no longer just parlor tricks; they are evolving into highly capable assistants that can handle the rigorous demands of actual research.[1][2][8]

Despite these impressive benchmark victories, a fundamental hurdle has historically prevented mathematicians from fully trusting AI: the problem of hallucination. In advanced mathematics, a proof can easily span over 100 pages of dense, interconnected logic. If an AI model confabulates even a single line—misapplying a theorem or dropping a crucial negative sign—the entire mathematical argument collapses. Because large language models are inherently probabilistic, they have traditionally been viewed as too unreliable for the absolute certainty required by the discipline.[5]

The solution to this trust deficit has not come from making language models less probabilistic, but by pairing them with an entirely different kind of software: the interactive theorem prover. The most prominent of these is Lean 4, a programming language and proof assistant that is rapidly becoming the gold standard in the mathematical community. Lean does not guess or estimate; it operates on strict, axiomatic logic, checking every single step of a mathematical argument against the foundational rules of mathematics.[4][5]

In this new paradigm, Lean acts as an "immutable verification oracle." When an AI system proposes a proof, it must write that proof in the Lean programming language. The Lean compiler then checks the code. If the code compiles successfully, the mathematics is guaranteed to be 100 percent correct. This architecture completely neutralizes the threat of AI hallucinations. The language model is free to be creative, intuitive, and even make mistakes, because the Lean kernel will instantly catch and reject any logical flaws before they are accepted as truth.[4][7]

This synthesis of generative AI and formal verification has given rise to the era of the "AI Co-Mathematician." Instead of functioning as autonomous oracles that spit out finished papers, modern AI systems act as tireless research assistants. They can rapidly search through vast databases of known theorems, propose multiple tactical approaches to a stubborn lemma, and grind through the tedious algebraic manipulations that often bog down human researchers. The human mathematician acts as the director, setting the high-level strategy while the AI handles the tactical execution.[4][6]

The real-world impact of this workflow is already transforming how landmark mathematics is done. In a widely celebrated milestone, a team led by Fields Medalist Terence Tao used Lean to formally verify the proof of the polynomial Freiman-Ruzsa conjecture. What would traditionally have been a painstaking, months-long process of peer review and manual checking was completed by a 25-person collaborative team in just three weeks. The formalization proved that complex, modern research results can be verified almost as quickly as they are written.[5]

This rapid acceleration is being fueled by the massive expansion of Mathlib, Lean's central library of formalized mathematics. Built by a global community of volunteer mathematicians and computer scientists, Mathlib has grown to encompass approximately 120,000 formalized lemmas as of early 2026. This repository functions as a comprehensive, machine-readable map of known mathematics. When an AI system like ByteDance's Seed Prover or OpenAI's latest models attempt to solve a new problem, they can autonomously search Mathlib to find and apply the exact foundational theorems required for their proofs.[5][7]

Armed with these tools, AI systems are beginning to chip away at longstanding open questions. Recent reports indicate that advanced models, working in tandem with formal verification environments, have contributed to solutions for several open Erdős problems—a famous set of mathematical conjectures posed by the legendary Paul Erdős. By breaking these problems down into smaller, verifiable steps, AI models are demonstrating that they can make novel contributions to the field with minimal human guidance, provided their output is anchored by a formal verifier.[4][6]

Yet, for all their tactical brilliance, AI systems still suffer from what researchers call the "definition-finding gap." While an AI can brilliantly execute a proof once a problem is clearly defined, it struggles immensely with the conceptual framing required to invent new mathematics. Human mathematicians rely on taste, intuition, and a deep sense of aesthetic beauty to decide which problems are actually worth solving. AI cannot yet look at a chaotic mathematical landscape and identify the hidden structures that warrant a new definition or a groundbreaking new theory.[1][2]

Consequently, the culture of mathematics is undergoing a profound shift. The romanticized stereotype of the lone genius working in isolation with a chalkboard is giving way to massive, digitally connected collaborations. Proof assistants and AI tools are democratizing the field, allowing researchers to build upon each other's verified work with absolute confidence. As the barrier to verifying complex proofs drops, mathematicians are free to tackle increasingly ambitious and sprawling conjectures that would have been impossible to verify by hand.[1][5]

Ultimately, the results of the First Proof benchmark and the rise of the Lean theorem prover point to an uplifting future for the discipline. Mathematics is not being "solved" or automated away by machines; rather, it is being upgraded. By offloading the burden of rigorous, line-by-line verification to AI and formal proof assistants, human mathematicians are being freed to focus on what they do best: exercising creativity, following their intuition, and dreaming up the next great questions to ask the universe.[1][2]