How AI and a Strict Coding Language Are Unlocking Decades-Old Math Mysteries

Jared Duker Lichtman spent four years of his doctorate proving that prime numbers are the most efficient sets of their kind, eventually publishing a celebrated proof of the Erdős Primitive Set Conjecture. He then spent the next seven years chasing the next open question in that same mathematical family. In early 2026, an artificial intelligence system solved that next step in approximately 80 minutes. The resulting proof was so elegant that Lichtman described it as a "Book Proof"—a reference to Paul Erdős's concept of a divine text containing the universe's most perfect mathematical solutions.[1]

This moment represents a historic inflection point. For years, artificial intelligence has steadily mastered high school mathematics and Olympiad-level competition problems. But in 2025 and 2026, the technology crossed a critical threshold, moving from solving known exercises to cracking open, research-level conjectures that have stumped the world's best minds for decades.[1][4]

The breakthroughs are arriving at a rapid pace. In May 2026, OpenAI announced that its reasoning models had successfully tackled the planar unit distance problem, a geometric puzzle first posed by Erdős 80 years ago. By drawing on diverse branches of mathematics, the AI discovered an entirely new family of geometric constructions that broke an assumption mathematicians had held for nearly a century.[2]

Simultaneously, Google DeepMind's AlphaProof Nexus system demonstrated the ability to autonomously navigate complex research literature, successfully proving nine open Erdős conjectures drawn from a formalized database. These milestones raise a fundamental question: why are large language models, which have historically been terrible at basic arithmetic, suddenly capable of profound mathematical discovery?[4]

The answer lies in a paradigm shift known as neuro-symbolic reasoning. Historically, large language models operate by predicting the next most likely word in a sequence. This probabilistic approach is excellent for writing poetry or generating code, but it is fatal in mathematics, where a single hallucinated variable or skipped logical step collapses an entire proof.[3]

To solve the hallucination problem, researchers paired the creative intuition of neural networks with the unforgiving logic of a strict referee. That referee is Lean, an open-source functional programming language and interactive theorem prover.[3][6]

In Lean, every mathematical definition, theorem, and proof must be written in mechanically verified code. The compiler acts as an absolute gatekeeper. It tracks the logical state of a proof step-by-step; if an AI proposes a logical leap that violates the rules of mathematics, the Lean compiler simply throws an error and refuses to compile the code.[4][6]

This pairing creates a powerful "data flywheel." The AI model acts as the creative engine, proposing possible tactical steps to solve a problem. Lean acts as the verifier. When the AI makes a mistake, Lean provides immediate, ground-truth error feedback, allowing the model to try again and iteratively self-correct without needing a human to grade its work.[3][4]

The results of this self-correction loop have been staggering. Researchers at Princeton University utilized this architecture to build Goedel-Prover-V2, an open-source system that achieved a 90% accuracy rate on the miniF2F mathematical benchmark. By relying heavily on Lean's feedback, the Princeton team achieved state-of-the-art results using models 80 times smaller than those deployed by commercial tech giants.[3]

Crucially, the AI is not merely brute-forcing its way through millions of possibilities. It is demonstrating genuine aesthetic elegance. In solving the von Mangoldt weight identity, the AI discovered a conceptual opening that human convention had completely overlooked, proving that machine reasoning can expand the stylistic boundaries of the discipline.[1]

Despite these triumphs, a core skepticism remains within the mathematical community: the problem of data contamination. Because modern AI models are trained on vast swaths of the internet, it is incredibly difficult to prove that a model is genuinely reasoning through a novel problem rather than simply regurgitating a similar proof it memorized during training.[5]

To definitively test the limits of machine reasoning, 11 leading mathematicians launched the "First Proof" initiative in February 2026. The researchers selected ten unpublished lemmas—minor component proofs—from their own active, ongoing research. Because these problems had never been posted online, they were guaranteed to be absent from any AI's training data.[5]

The mathematicians encrypted the answers and gave AI systems a one-week window to attempt the problems. The preliminary results offered a sobering reality check: while AI systems are incredibly powerful, they still struggle to autonomously clear the board on entirely novel, unguided deep research without human prompting.[5]

Rather than replacing mathematicians, AI is rapidly becoming an indispensable co-pilot. Working researchers are beginning to offload the tedious verification of minor lemmas to AI systems, freeing their own cognitive bandwidth for high-level architectural thinking and the formulation of new conjectures.[4][5]

This collaborative dynamic is democratizing the field. With Lean and AI assistants, students and researchers can verify their own proofs instantly, lowering the barrier to entry for rigorous mathematics. As the fusion of human intuition and formal verification continues to mature, mathematics is entering a golden age of accelerated discovery.[3][6]