This post analyzes AlphaProof Nexus: A System for Autonomous Mathematical Discovery, by the Google DeepMind team (arXiv:2605.22763).
The Core Problem: Proofs vs. AI Hallucination
LLMs can produce mathematically fluent arguments containing subtle logical breaks that even top human reviewers may miss within hours. AlphaProof Nexus's answer is radical: let the compiler review, not humans.
Core Mechanism: LLM Intuition + Lean's Rigor
- Gemini 3.1 Pro proposes proof strategies and drafts proofs, like an intuitive mathematician.
- Lean 4 compiler verifies every step against formal rules. Any invalid step is rejected immediately.
- Erdős problems (353 open): autonomously solved 9, including 2 that had been open for 56 years (posed in 1970)
- OEIS conjectures (492 open): proved 44
- Solved a 15-year-old Hilbert function problem; improved known bounds in convex optimization
- Cost: a few hundred dollars in inference per problem
- OpenAI route: GPT-5.5 Pro outputs natural-language proofs. Fields Medalist Gowers fed it the Erdős unit distance conjecture and received PhD-thesis-level results in two hours, overturning an 80-year-old conjecture—but verification requires expert human review.
- DeepMind route: Lean formal proofs, automatically compiler-verified. No subjective judgment; right is right, wrong is wrong.
- Limited Lean library coverage: algebraic geometry and topology lag far behind number theory and combinatorics.
- Long iteration times: hard proofs may take days to weeks, versus a mathematician's flash of insight.
- Hardest problems remain unsolved: only 9 of 353 Erdős problems (~2.5% success rate).
- Paper: https://arxiv.org/abs/2605.22763
- Source: Google DeepMind
The loop runs continuously: generate → verify → on error, feed back compiler messages → revise → re-verify. A proof is accepted only when it compiles.
Key insight: the compiler feedback loop is the soul of the system. Ablation experiments show that even the simplest agent (Gemini + Lean basic loop, no evolutionary algorithm, no AlphaProof assistance) can still prove the 9 solved Erdős problems. Complex modules are enhancements, not prerequisites.
Four-Tier Agent Architecture
| Tier | Capability | Key difference | |------|-----------|----------------| | Agent A | Gemini + Lean basic loop | Multi-turn dialogue to revise Lean code with real-time compiler feedback | | Agent B | + AlphaProof access | Fills missing proof fragments, calls existing proof tools | | Agent C | + evolutionary mechanism | Shares, scores, and ranks multiple proof drafts (similar to AlphaEvolve) | | Agent D | full integration | Most complete version, for the hardest problems |
Counter-intuitively, the 9 solved Erdős problems only required Agent A—just at higher cost. The core mechanism is strong; elaborate architecture is efficiency optimization, not a necessity.
Results
Why This Matters
1. Zero-error guarantee: every proof is formally verified by Lean. No "looks right but isn't" gray zone. 2. Beyond retrieval: these Erdős problems had no known standard solutions; the AI explored, tried, and corrected in genuinely unknown territory. 3. Stunningly cheap: hundreds of dollars to crack a world-class open problem. 4. New workflow for mathematicians: even incomplete formal proof drafts are valuable—mathematicians can focus directly on unresolved subgoals without re-verifying the whole chain.
Two Rival Routes (May 2026)
| | OpenAI route | DeepMind route | |-|-|-| | Output | Natural-language proofs | Lean formal code | | Verification | Human expert line-by-line review | Automatic compiler check | | Flexibility | High, not limited by formalization libraries | Limited to math formalized in Lean | | Certainty | Depends on human judgment | 100% machine-verified | | Speed | Fast (two hours) | Slow (iterative) | | Cost | Low (single inference) | Medium (hundreds of dollars/problem) |
The routes are complementary, not exclusive: natural language for exploring ideas, formal systems for locking down verification.
Open Challenges
Still, as the Lean math library grows and models improve, that rate should climb.
Conclusion
Paul Erdős left 1,217 prize-backed problems behind. He likely never imagined the claimant would not be human. AlphaProof Nexus shows that when LLM creativity meets formal constraint, AI can make verifiable progress in humanity's most abstract and rigorous intellectual activity. This is not the end of mathematics—it is the beginning of a new era.