AlphaProof Nexus Cracks Decades-Old Math Problems for Hundreds of Dollars
Summary
- • Google DeepMind's AlphaProof Nexus solved 9 of 353 open Erdős problems, including two unsolved for 56 years
- • System proved 44 of 492 OEIS conjectures and settled a 15-year-old algebraic geometry question
- • Inference cost just a few hundred dollars per problem using Gemini 3.1 Pro with Lean compiler feedback
- • Simplest agent variant replicated all 9 solutions — signaling shift from specialized systems toward simple agentic loops
Details
AlphaProof Nexus solved 9 of 353 open Erdős problems, including two unsolved for 56 years
The Erdős problems are a canonical benchmark of open combinatorics and number theory questions posed by prolific mathematician Paul Erdős, many carrying cash prizes and decades of failed attempts by professional mathematicians. Solving 9 autonomously — even a small fraction — represents a meaningful threshold for AI-assisted mathematical research.
System proved 44 of 492 open OEIS conjectures, settled a 15-year-old algebraic geometry question, and improved a convex optimization bound
The breadth of results across multiple mathematical domains — integer sequences, algebraic geometry, and optimization — demonstrates that the framework is not narrowly tuned to a single problem class. Improving a known bound in convex optimization has direct relevance to machine learning theory and algorithm design.
Core architecture pairs Gemini 3.1 Pro proof generation with Lean compiler verification in a closed error-feedback loop
By generating proof steps in Lean rather than natural language, the system gets mechanically verified feedback after each step. Compiler error messages are fed directly back to the model as grounding signals. This offsets LLMs' well-documented weakness in sustained logical reasoning without requiring fine-tuning on mathematical data for every new problem class.
Four agent variants scale from simple Gemini-Lean loops (Agent A) to full evolutionary search with AlphaProof integration (Agent D)
Agent B adds AlphaProof RL queries to fill proof gaps. Agent C adds an evolutionary population of proof sketches scored by Gemini 3.0 Flash via an Elo system. Agent D combines all capabilities. Post-hoc analysis found Agent A alone could reproduce all nine Erdős solutions — suggesting the complex scaffolding may become redundant as LLMs improve.
DeepMind researchers describe an ongoing shift from specialized trained systems toward simple agentic loops as LLMs improve
This framing positions AlphaProof Nexus not as a fixed product but as evidence of a broader architectural trend: as general-purpose LLMs grow more capable, the value of domain-specific training pipelines and complex multi-agent scaffolding may decline. Researchers attribute Agent A's competitive performance to 'rapid improvement in underlying LLMs and the power of compiler feedback.'
System's successes concentrate where Lean's Mathlib library is mature; problems requiring new mathematical theory remain out of reach
AlphaProof Nexus performs best in combinatorics, number theory, and convex optimization — areas with extensive Lean formalization. The 344 unsolved Erdős problems and 448 unproven OEIS conjectures highlight that the system cannot yet synthesize fundamentally new mathematical frameworks, where most of the frontier of pure mathematics actually lives.
Inference cost of a few hundred dollars per problem sets a new economic benchmark for AI-assisted mathematical research
At this price point, research institutions and individual mathematicians could realistically run the system on open problems without large compute budgets. This democratizes access to AI proof assistance and raises the question of whether formal verification infrastructure — and Lean expertise — becomes a new bottleneck for mathematical AI adoption.
Research = experimental results and findings, Tech Info = system architecture and mechanisms, Insight = researcher-attributed trend claim, Other = scope constraints and limitations, Market Impact = cost and adoption implications
What This Means
AlphaProof Nexus marks a concrete step toward AI systems that can contribute to — not merely assist with — genuine mathematical research, resolving specific open problems that have resisted human effort for decades. The low inference cost and the finding that simpler agent configurations perform comparably to complex ones as base LLMs improve suggest this capability will become more accessible and more powerful without requiring proportional increases in engineering complexity. For practitioners and investors, the more significant signal may be architectural: if formal verification loops can reliably ground LLM reasoning in hard mathematical domains, the same pattern is likely to propagate into scientific reasoning, theorem-driven software verification, and any domain where a compiler or symbolic checker can serve as an objective arbiter.
Sentiment
Mostly excited about formal verification loops and low-cost breakthroughs, with some practical questions
“Google DeepMind just solved 9 math problems that stumped humans for 56 years... The compute cost to solve a half-century-old math problem was just a few hundred dollars. The AI doesn't need to be perfect on the first try. It just needs a flawless feedback loop and enough time to think.”
“DeepMind's AlphaProof Nexus paper... The AI generates proofs in Lean... If ANY step is wrong, the proof is rejected. Translation: The AI can't fake it. There's no hallucination problem.”
“quick take on AlphaProof Nexus: this isn’t just “AI solved math”. it’s a real agentic loop (LLM + formal verifier) that can actually prove things. this is the kind of primitive i want to plug into my own agents.”
“AlphaProof Nexus combines LLMs with the Lean proof assistant to automatically discover and verify formal proofs, cracking several open math problems. A striking step toward AI-driven mathematics.”
Split
~80/20 positive/neutral — strong agreement on the significance of the Lean-verified agentic loop; minor questions on efficiency comparisons to existing tools like Lean Copilot.