axiommath.aiAxiom
A mathematical reasoning engine building provably correct AI starting with an AI mathematician
- Stage
- Series A
- Sector
- AI / Formal Reasoning
- ROUND
- $200M
- Post
- $1.6B
- Team
- 35 FTEs · Palo Alto
Axiom uses formal mathematical language to generate machine-checkable, step-by-step proofs, the opposite of pattern-matching LLMs. In under six months, the team posted the best public results in AI-for-math (Putnam 12/12, two Erdős problems), assembled arguably the densest technical team in the field, and opened commercial conversations with NVIDIA on chip verification.
First AI ever to solve every problem on the world's hardest undergrad math competition. Only 5 perfect human scores in 98 years.
Formal Lean proofs of problems open for 45 and ~30 years.
World's foremost code verification benchmark. A 28× improvement over the prior state-of-the-art reported in that paper.
The Thesis
LLMs generate plausible-sounding answers via pattern matching yet they cannot verify their own reasoning. In domains where "probably right" isn't good enough (chip design, cryptography, aerospace, quant), verifiability is paramount. Axiom's insight is that the language of correctness is formal mathematical proof, not next-token prediction.
By training systems to output formally verified proofs, every step is machine-checkable. The system knows when it's wrong (the proof won't compile), and every verified proof becomes new training data.
Every conclusion has a proof chain back to axioms.
The system knows when it's wrong because the proof doesn't compile.
Same reasoning infra applies across STEM domains.