đŸ”¬Scaling Past Informal AI - Carina Hong, Axiom Math

· Source: Latent.Space - Www.latent.space · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences, Software Development & Engineering · Depth: Advanced, extended

Summary

Axiom Math, led by CEO Carina Hong, recently secured a \$200 million Series A funding round, valuing the company at \$1.6 billion. Axiom specializes in "verified AI" and formal verification for mathematics, leveraging the Lean proof assistant. The company achieved a perfect score of 120 points on the December 2025 Putnam exam, surpassing top human competitors (110 points) and leading LLMs like DeepSeek (103 points). Axiom's methodology involves post-training LLMs on Lean data using reinforcement learning and supervised fine-tuning, recursively decomposing proof goals. They are also open-sourcing mathematical discovery tools for pre-conjecturing and construction finding. Axiom recently launched Axle, an API offering Lean proof validation and manipulation tools, which is 100 times faster than Comparator. The company posits that verified AI scales brilliance and provides significant performance gains, extending its market vision to critical hardware and software verification.

Key takeaway

For Directors of AI/ML evaluating reasoning capabilities for critical systems, recognize that formal verification, exemplified by Axiom Math's achievements, is a strategic imperative, not merely a compliance burden. Your teams should explore integrating formal proof assistants like Lean and tools such as Axle to achieve superhuman performance, enhance sample efficiency, and ensure verifiable AI outputs, fundamentally shifting how you approach complex problem-solving and system reliability.

Key insights

Verified AI, grounded in formal mathematics, scales human brilliance and offers significant performance advantages over informal methods.

Principles

Method

Axiom's system post-trains LLMs on Lean proof data via RL/SFT, recursively decomposing proof goals into sub-goals and incorporating backtracking, augmented by meta-programming tools for validation.

In practice

Topics

Best for: AI Scientist, Director of AI/ML, Machine Learning Engineer

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Latent.Space - Www.latent.space.