Artificial intelligence's recent leap into advanced mathematics, including solving long-open problems in pure math and theoretical computer science, didn't require a new kind of intelligence at all. The same training recipe that powers coding assistants now solves problems mathematicians have puzzled over for decades. But AI's path to mathematical prowess took a detour compared to its rapid progress in software development, and understanding why reveals how AI systems actually improve.

What's the Real Difference Between AI's Math and Coding Breakthroughs?

The mechanism behind both achievements is identical: scale up the model, feed it massive amounts of training data, apply reinforcement learning with verifiable rewards (RLVR), and let the system learn from feedback on whether its solutions work. In May 2026, an OpenAI model disproved the Erdős unit-distance conjecture, a problem in geometry that had stumped mathematicians for years. In August 2026, an internal version of OpenAI's next model family, called Astra, announced ten results on long-open problems in mathematics and theoretical computer science.

Yet coding assistance became commercially viable years earlier. The difference wasn't intelligence or capability; it was infrastructure. Here's how the training process actually works: the model generates candidate solutions step by step, a surrounding system called a harness fetches the data the model needs, verifies whether the solution is correct, and returns feedback so the model can improve. A reward engine then assigns scores (for example, 1 if a proof checks out, 0 if it doesn't), and reinforcement learning adjusts the model's behavior based on those rewards.

Why Did Mathematics Take So Long to Catch Up?

Three concrete obstacles delayed AI's mathematical breakthroughs compared to its progress in coding:

• Training Data Scarcity: Coding had GitHub, containing hundreds of millions of repositories of working, reviewed, machine-readable code. Mathematics has no equivalent. The largest library of formalized mathematics, Lean's mathlib, contains roughly two million lines of code accumulated over nine years, a fraction of the available code corpus. Informal mathematical writing is additionally scarce, dense, and often paywalled.
• Missing Verification Infrastructure: Reinforcement learning requires a cheap, reliable judge of completed attempts. Coding had compilers, test suites, and continuous integration pipelines from the start. Mathematics only recently acquired its equivalent in the form of formal proof verification systems.
• Weak Economic Incentives: Labs allocate effort in rough proportion to expected return. Coding assistance is a market worth hundreds of billions of dollars; pure mathematics has little direct commercial value. This alone predicts that coding would be industrialized first.

The verification infrastructure timeline tells the story clearly. In October 2025, OpenAI researchers claimed GPT-5 had solved previously unsolved Erdős problems, but within days it became clear the model had simply retrieved existing papers. There was no verification step at all. In May 2026, the unit-distance disproof was validated by a panel of external mathematicians, but verification was human and slow. By August 2026, the ten Astra results shipped with Lean 4 formalizations, machine-checkable proofs with no unfinished goals and only standard axioms, published openly alongside the manuscripts.

"To this end, we developed a framework, AlphaProof Nexus, for LLM-aided proof generation and used it to build a basic agent in which a set of subagents independently searches for proofs with feedback from the Lean compiler," according to DeepMind researchers describing their May 2026 system.

DeepMind Research Team, AlphaProof Nexus Paper

The August 2026 release shows how seriously the verification layer is now engineered. The published repository pins the toolchain (Lean 4.32.0, mathlib) and includes "comparator challenges": instructions for re-checking each proof by compiling it in a sandbox, exporting the proof terms, and replaying them through both Lean's own kernel and an independently implemented external checker written in Rust, while confirming that the proved statement is identical to the claimed theorem and that no placeholder or hidden axiom was used.

How to Understand the Role of Verification in AI Training

• Reward Signal Creation: A reliable way to score completed attempts is essential. In coding, compilers and test suites provide instant feedback. In mathematics, formal proof checkers now serve this role, allowing systems to know immediately whether a proposed solution is correct.
• Feedback Loop Acceleration: Once verification is automated and cheap, reinforcement learning can run at scale. The model generates thousands of candidate solutions, the verification system scores them, and the model learns which approaches lead to correct answers, dramatically speeding up improvement.
• Trust and Reproducibility: Automated verification creates a gold standard for validating proofs from untrusted sources. The Lean community's documented verification stack was explicitly designed for scenarios like "unaligned AI," meaning the infrastructure was built in anticipation of machine-generated mathematics before it became routine.

DeepMind's AlphaProof Nexus system, published in May 2026, solved 9 Erdős problems out of 353 attempted, including two questions that had been open for 56 years, at an inference cost of a few hundred dollars per problem. It also proved 44 open conjectures from the Online Encyclopedia of Integer Sequences, resolved a 15-year-old open question on Hilbert functions in algebraic geometry, and improved an open bound in convex optimization by discovering a novel algorithmic parameter schedule.

What comes next is already visible. Big labs are ramping up their hiring of top talent in mathematics and science, signaling that the speed of progress will accelerate further. The bottleneck was never raw intelligence; it was data, verification tools, and economic motivation. Now that all three are in place, the pace of AI-assisted mathematical discovery is likely to become routine rather than exceptional.