AI Systems Solve Decades-Old Erdős Math Problems, Sparking a New Era of Formal Verification

For decades, artificial intelligence has served mathematics primarily as a high-speed calculator, crunching numbers while humans provided the structural insight. In late May 2026, that dynamic fundamentally shifted. Two of the world's leading AI laboratories announced back-to-back breakthroughs in which neural networks produced genuinely novel, research-level mathematical discoveries.[1]

The first milestone arrived when OpenAI researchers revealed that their model had solved the "unit distance problem," a notorious combinatorial geometry puzzle proposed nearly 80 years ago by the legendary Hungarian mathematician Paul Erdős. Rather than validating the long-standing conjecture, the AI generated a complex counterexample—constructing a point arrangement with more unit-distance pairs than the accepted upper bound allowed.[1][2]

Just a day later, Google DeepMind unveiled AlphaProof Nexus, an AI system that autonomously solved nine open Erdős problems. These problems, spanning graph theory and combinatorics, included two that had remained unsolved for 56 years. DeepMind achieved this at a compute cost of only a few hundred dollars per problem, while simultaneously proving 44 open conjectures from the Online Encyclopedia of Integer Sequences.

The engine driving this sudden acceleration is not just a larger language model, but the marriage of generative AI with formal proof assistants. Large language models (LLMs) are highly capable of proposing creative mathematical steps, but they are prone to logical hallucinations. To solve this, researchers paired them with Lean, a rigorous programming language and theorem prover.

Lean acts as an uncompromising referee. When the LLM generates a potential proof step, Lean checks its logical validity against foundational axioms. If the step is flawed, the system rejects it, forcing the AI to try a new path. This iterative feedback loop—propose, verify, repeat—allows the AI to explore vast mathematical spaces without drifting into error.[4]

The infrastructure supporting this neuro-symbolic approach has been quietly building for years. Lean’s collaborative mathematical library, mathlib4, expanded rapidly through community effort, surpassing 250,000 formalized theorems and 120,000 definitions by late 2025. This massive repository of machine-readable mathematics provides the training ground and the building blocks for systems like AlphaProof Nexus.

The transition of AI from a parlor trick to a foundational research tool is now attracting major institutional backing. In May 2026, the Defense Advanced Research Projects Agency (DARPA) awarded a $2.6 million grant to researchers at UC Irvine and USC. The three-year project is designed to measure exactly how AI tools accelerate progress when deployed in active, real-world mathematical research environments.[3]

Rather than testing AI on closed problem sets with known answers, the DARPA-funded initiative convenes expert mathematicians to tackle unsolved problems in number theory and partial differential equations alongside AI co-pilots. The goal is to quantify AI's impact on genuine mathematical discovery, moving beyond simple problem-solving accuracy to evaluate its role in frontier research.[3]

Despite the undeniable results, the rise of AI-generated proofs has sparked a fierce debate over the philosophical purpose of mathematics. The core tension lies in the difference between knowing that a theorem is true and understanding why it is true. Proofs that are verifiable in Lean are not always parse-able by human minds.

David Bessis, a mathematician and science writer, has voiced concerns that autoformalization and AI-generated proofs could reduce the conceptual benefit the mathematics community gains from new discoveries. If an AI solves an 80-year-old problem using a convoluted, million-step logical derivation that no human can follow, the field gains a factual answer but no new intuition or theoretical framework.

Proponents counter that this is a temporary bottleneck. Researchers are actively developing bidirectional translation tools to convert dense Lean code back into natural-language sketches and lemmas that humans can engage with. Furthermore, the Lean Focused Research Organization (FRO), established in 2023, is explicitly tasked with improving Lean's usability, documentation, and proof automation to bridge the gap between human mathematicians and machine verification.[4]

The implications extend far beyond abstract geometry. The ability to formally verify complex logic at scale has immediate applications in software and hardware verification, ensuring that critical systems—from cloud infrastructure to medical devices—operate without catastrophic bugs. The same neuro-symbolic architecture solving Erdős problems is already being adapted to verify AWS security policies and optimize numerical algorithms.[4]

The cultural shift is already visible in how the next generation of mathematicians is being trained. In early 2026, the International Centre for Mathematical Sciences (ICMS) hosted a dedicated residency on AI and mathematics, training students in the emerging pipeline of conjecture generation, autoformalization, and automated theorem proving. Participants explored how modern neuro-symbolic systems blend deep learning with formal reasoning, signaling that fluency in AI tools is becoming as essential as fluency in calculus.

Mathematics is undergoing a structural phase transition. The May 2026 breakthroughs demonstrate that AI is no longer just an assistant for tedious calculations; it is an autonomous engine for discovery. As the mathematics community navigates the tension between machine-verified truth and human comprehension, the discipline is expanding into territories that neither humans nor machines could explore alone.[1][2]