A new wave of artificial intelligence has pushed into territory scientists once thought of as off-limits. According to Wired.com, an AI startup named Axiom has developed a system called AxiomProver that recently produced solutions to four mathematical problems that had remained unsolved despite years of work by experts.
The breakthrough came after mathematician Dawei Chen and his colleague Quentin Gendron left one problem as a conjecture because traditional methods failed to deliver a proof. At a conference, Chen encountered Axiom co-founder Ken Ono, who used AxiomProver to derive a proof the next day. The system identified a connection to a 19th-century numerical phenomenon that had eluded human researchers and verified its result using the formal Lean proof language, a tool for ensuring mathematical rigor.
That proof is just one piece of Axiom’s recent success. Among the four problems the AI tackled is Fel’s Conjecture, an algebraic geometry challenge tied to formulas from Indian mathematician Srinivasa Ramanujan’s notebook. In that case, AxiomProver didn’t just fill a gap in a human attempt; it constructed and verified the complete proof autonomously. Other victories by the system span areas such as probability in number theory and techniques originally developed around the famed Fermat’s Last Theorem.
What sets AxiomProver apart from standard large language models is its combination of generative reasoning and formal verification. Rather than relying only on pattern matching from training data, the engine can construct proofs that meet the exacting standards of professional mathematics. Researchers involved describe this as a shift in how AI can augment scientific discovery rather than replace mathematicians, offering a tool that expands what’s tractable in research.
Beyond pure mathematics, the techniques behind AxiomProver could influence fields that depend on provable correctness, such as cybersecurity and software verification. For now, the achievement underscores a broader trend: AI systems are inching toward genuine reasoning in domains once considered uniquely human.