Google DeepMind's AlphaProof Nexus Solves Open Math Problems

What changed

Google DeepMind announced that its AlphaProof Nexus framework has autonomously solved dozens of long-standing open mathematical problems. The system cracked nine of 353 open Erdős problems it attempted, including two that remained unsolved for 56 years. In addition, AlphaProof Nexus proved 44 conjectures from the Online Encyclopedia of Integer Sequences (OEIS), resolved a 15-year-old question about Hilbert functions in algebraic geometry, and improved a known bound in convex optimization.

The work appears in a peer-reviewed study published in Science, with the formal proofs and supporting materials made available on GitHub. This approach contrasts with a recent OpenAI announcement of hundreds of AI math proofs posted directly to GitHub without journal submission. The Association for Human Mathematics, led by Fields Medalist Terence Tao, called for a boycott of this practice, citing concerns about bypassing traditional peer review.

The AlphaProof Nexus system is built around Google’s Gemini 3.1 Pro language model. Rather than generating complete proofs in natural language alone, the framework routes proof drafts through Lean, a formal language whose compiler verifies each logical step. When the compiler finds errors, it returns specific error messages that guide the next attempt. For harder subproblems, the system can call on AlphaProof, a specialized reinforcement learning tool designed for Olympiad mathematics. Notably, the simplest variant of the system solved all nine Erdős problems without needing this specialized module, though at higher cost for the toughest cases.

The framework uses four agent configurations, ranging from simple to complex. The basic agent loops between the language model and compiler feedback. More advanced versions add evolutionary selection of proof sketches, guided by rating agents that score candidates on plausibility and novelty using an Elo ranking system. The most sophisticated agent combines all these capabilities. The researchers found that improving language models and compiler feedback strengthens the simpler agent configurations over time, suggesting a trend toward less complex systems as underlying models improve. The inference cost per problem ran to a few hundred dollars.

Why it matters

This represents a concrete capability milestone for AI in pure mathematics research. The problems AlphaProof Nexus solved were genuinely open, confirmed by peer review, and required formal verification rather than relying solely on natural language reasoning. The system’s architecture offers a practical model for AI mathematics: pairing language model reasoning with symbolic verification to catch errors that models typically make in logical deduction.

For mathematicians and research institutions, the approach demonstrates that AI tools can contribute to real research workflows. Several mathematicians who worked with the system reported that even failed proof attempts deepened their understanding of problems. Formal proof sketches let experts focus on unsolved subgoals rather than re-checking entire arguments. The system also helped surface errors in formalized versions of published proofs, suggesting it can serve as a quality filter for mathematical literature.

The speed and low cost matter as well. Running specialized AI math research at a few hundred dollars per problem makes it accessible to researchers without massive computing budgets. The framework shows a trend toward simpler agent loops as underlying language models improve, potentially widening access further over time. The system is already being used in ongoing research on quantum optics and graph theory.

However, the success rate remains modest. The system solved only about 2.5 percent of the Erdős problems it tackled, concentrated in areas like combinatorics, number theory, and convex optimization where the Lean proof library is mature. Most open problems stayed out of reach, and problems requiring extensive new theory proved intractable. Terence Tao noted that AI success rates across Erdős problems hover around one to two percent overall, with successes confined to relatively easier questions. The system inherits the reliability issues of large language models, meaning inconsistency and hallucination remain concerns even with symbolic verification.

What to test

Before integrating AlphaProof Nexus into a research workflow, organizations should verify several claims and practical limitations:

• Confirm the peer-review standards applied to the published study and whether reviewers had access to code and data for independent verification.
• Test whether the system performs reliably across problem domains beyond combinatorics, number theory, and convex optimization. The source notes that Lean’s math library is not yet mature in all areas.
• Assess how often the system produces formally correct but trivial proofs, or proofs that sidestep the core difficulty of a problem by relying on unproven lemmas.
• Measure actual inference costs on your own problems and hardware to confirm the few-hundred-dollar estimate applies to your use case.
• Evaluate whether failed proof attempts actually improve researcher intuition as claimed, or whether incorrect reasoning from the model leads researchers astray.
• Determine if Lean proof artifacts are sufficiently detailed for human mathematicians to identify and focus on remaining open subgoals, or if they remain opaque.
• Test whether the system can detect errors in published proofs at a useful rate, or if false positives waste reviewer time.

The conclusion

AlphaProof Nexus shows that formal verification combined with language model reasoning can tackle genuine open problems in mathematics. The peer-reviewed publication and public release of proofs on GitHub represent a responsible approach to AI math research and signal commitment to producing reliable, auditable results.

That said, the 2.5 percent success rate and concentration in specific domains clarify the limits. This is not a tool for solving arbitrary hard problems. It works best on questions that decompose into manageable subgoals and exist in domains where formal proof libraries are well-developed. The system strengthens human mathematicians rather than replacing them, particularly through its ability to offer structured partial results and catch errors in formalization.

Watch for three developments. First, whether the success rate rises as underlying language models improve and Lean libraries expand to new areas. Second, how widely researchers adopt the system for collaborative proof development, and whether that changes how mathematics research is organized. Third, whether formal verification approaches like this become standard practice in AI math research, or whether the field continues to debate the merits of natural language reasoning alone.

Related reading

• Sakana AI MLR Catches 73% of Core-Claim Errors in Papers
• EmbeddingGemma 2: Google’s 740M On-Device Embedding Model
• Ai2 Replaces Priority Scheduling With GPU Time Budgets