> Paper: AlphaProof Nexus: A System for Autonomous Mathematical Discovery > Authors: Google DeepMind team > arXiv: https://arxiv.org/abs/2605.22763
The Core Problem: Mathematical Proof vs. AI Hallucination
Mathematicians have long feared proofs that look correct but hide subtle flaws. Large language models amplify this nightmare: they can produce fluent, symbol-perfect arguments containing invisible logical gaps 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 Iron Discipline
The system couples two components:
- Gemini 3.1 Pro — proposes proof strategies like an intuitive mathematician, generating draft proofs as Lean code.
- Lean 4 compiler — rigorously verifies every step. It does not read "vibes," only logic; any invalid derivation is rejected immediately.
- Erdős problems (353 open): autonomously solved 9, including 2 that had stood for 56 years (posed in 1970)
- OEIS conjectures (492 open): proved 44
- Other breakthroughs: resolved 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 line-by-line expert review.
- DeepMind route: AlphaProof Nexus outputs Lean formalized proofs, machine-verified. No subjective judgment, no review disputes—right is right, wrong is wrong.
- Limited Lean library coverage: algebraic geometry, topology, and related fields are far less formalized than number theory and combinatorics, so the system cannot yet reach them.
- Long iteration times: complex proofs may take days or weeks, versus a mathematician's flash of insight.
- Hardest problems remain unsolved: only 9 of 353 Erdős problems were solved—a roughly 2.5% success rate.
- Paper: https://arxiv.org/abs/2605.22763
- Source: Google DeepMind
The loop runs repeatedly: Gemini generates → Lean verifies → errors are fed back → Gemini revises → re-verify. A proof is accepted only when the compiler passes it.
Key insight: the compiler feedback loop is the system's soul. Ablation experiments show that removing the evolutionary algorithm and AlphaProof assistance—keeping only the basic Gemini + Lean loop (Agent A)—still suffices to prove the 9 solved Erdős problems. Other modules are icing; the compiler feedback loop is the lifeline.
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 integration | Fills in 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 | Complete version for the hardest problems |
Counterintuitively, while the hardest problems require Agent D, the 9 solved Erdős problems can also be solved by Agent A—at higher cost. The core mechanism is strong enough; the complex architecture is an efficiency optimization, not a prerequisite.
Results
Why It Matters
1. Zero error, guaranteed
Every proof passes formal verification by the Lean compiler. There is no gray zone of "looks right but might be wrong." Compiler-passing means the proof is correct—an assurance natural-language proofs can never provide.2. AI beyond "retrieval mode"
These Erdős problems have no known standard solutions. The AI was not matching templates from a database but exploring, trying, and correcting in unknown territory, finding paths humans missed for half a century.3. Shockingly low cost
A few hundred dollars per world-class mathematical problem—nearly free on the scale of mathematicians' research budgets.4. Changing how mathematicians work
Even when AI cannot fully prove a target theorem, its formalized proof attempts have research value: mathematicians can focus directly on the unresolved subgoals without re-verifying the whole argument chain. AI does not replace mathematicians—it frees their attention from tedious verification toward the gaps that truly need human intuition.Two Rival Routes
In May 2026, two very different routes in AI mathematics broke through simultaneously:
| | OpenAI route | DeepMind route | |-|-|-| | Output | Natural-language proofs | Lean formalized code | | Verification | Human expert review | Automatic compiler checks | | Flexibility | High, unconstrained by formal libraries | Limited to math formalized in Lean | | Certainty | Depends on human judgment | 100% machine-verified | | Speed | Fast (results in two hours) | Slow (iterative) | | Cost | Low (single inference) | Medium (hundreds of dollars per problem) |
The two routes are complementary, not mutually exclusive. The most likely future: natural language for exploring ideas, formal systems for solidifying verification. Human mathematicians pose problems, steer direction, and distill insights; AI searches vast possibility spaces and verifies rigorously.
Open Challenges
But the trend is clear: as the Lean math library expands and model capability improves, this success rate will only rise.
Conclusion
Paul Erdős left behind 1,217 prized mathematical problems. He may never have imagined that one day the prize would be claimed not by a human, but by a machine. AlphaProof Nexus proves that when the creative intuition of an LLM is combined with the strict constraints of a formal system, AI can make verifiable progress in humanity's most abstract and rigorous intellectual activity.
This is not the end of mathematics, but the beginning of a new era.