TLDR: AlphaEvolve is a new AI system that combines large language models with evolutionary computation to autonomously discover mathematical constructions and improve solutions to open problems. It was tested on 67 diverse mathematical problems, achieving new bounds and generalizing results. The system operates in ‘search’ and ‘generalizer’ modes, evolving programs to find or create optimal solutions. It also integrates with other AI tools for proof generation and formal verification, demonstrating a powerful human-AI collaboration for advancing mathematical discovery.
A groundbreaking new research paper titled “Mathematical Exploration and Discovery at Scale” introduces AlphaEvolve, an advanced artificial intelligence system designed to autonomously discover novel mathematical constructions and push the boundaries of long-standing open problems. Authored by Bogdan Georgiev, Javier Gómez-Serrano, Terence Tao, and Adam Zsolt Wagner, this work showcases a significant leap in AI’s capability to contribute to mathematical research.
AlphaEvolve is described as a generic evolutionary coding agent that ingeniously combines the generative power of large language models (LLMs) with automated evaluation within an iterative evolutionary framework. This system proposes, tests, and refines algorithmic solutions to complex scientific and practical problems, demonstrating a new paradigm for mathematical discovery.
To illustrate its broad applicability, the researchers tested AlphaEvolve on a diverse list of 67 problems spanning various fields of mathematics, including analysis, combinatorics, geometry, and number theory. The system proved remarkably effective, rediscovering the best-known solutions in the majority of cases and, more impressively, discovering improved solutions in several instances. In some scenarios, AlphaEvolve even demonstrated the ability to generalize results from a finite number of input values into universal formulas.
The core innovation behind AlphaEvolve lies in its evolutionary approach. Unlike traditional methods that might search directly for mathematical objects, AlphaEvolve evolves Python programs that either directly generate constructions or, more commonly, act as search heuristics to find constructions. This ‘search mode’ allows the system to leverage expensive LLM calls to design efficient search strategies, which can then execute massive, cheap computations to explore millions of candidate constructions. This dynamic process allows for early-stage heuristics to make broad improvements, while later-stage heuristics fine-tune near-optimal configurations.
Furthermore, AlphaEvolve features a ‘generalizer mode’ where it is tasked with writing programs capable of solving problems for any given input value. By observing its own optimal solutions for smaller inputs, the system can identify patterns and generalize them into constructions valid for all inputs. This mode has led to some of the most exciting results, inspiring new research papers by human mathematicians.
The paper highlights AlphaEvolve’s ability to integrate with other specialized AI tools, creating a powerful pipeline for mathematical discovery and verification. For example, a construction discovered by AlphaEvolve for the finite field Kakeya problem was fed to Deep Think, another AI agent, which successfully derived a proof of its correctness and a closed-form formula. This proof was then formally verified in the Lean proof assistant using AlphaProof, demonstrating a seamless workflow from pattern discovery to formally verified mathematical results.
Among its many achievements, AlphaEvolve made notable improvements in several areas. It found lower-order improvements for Kakeya and Nikodym sets in dimensions 3, 4, and 5. It improved bounds for various autocorrelation inequalities, enhanced the upper bound for difference bases, and even raised the lower bound for the kissing number in 11 dimensions from 592 to 593. In packing problems, AlphaEvolve improved best-known results for packing hexagons and cubes, and for circle packing in squares and rectangles. It also found elusive configurations for subsets of grids with no isosceles triangles and improved lower bounds for the “no 5 on a sphere” problem. For the Ring Loading Problem, it improved the lower bound, and for the 3D Moving Sofa problem, it provided a construction with a higher estimated volume than previously known candidates. The system also successfully found optimal solutions for perfect square values of ‘n’ in a problem proposed for the International Mathematical Olympiad 2025.
The researchers emphasize the crucial role of human guidance in AlphaEvolve’s success. Careful design of evaluation functions (verifiers), the use of continuous loss functions, and expert advice in prompts significantly influence the system’s performance and the quality of discovered results. This suggests a future where human expertise and AI capabilities complement each other to solve complex mathematical challenges.
Also Read:
- New Benchmarks Uncover AI’s Challenges in Advanced Formal Algebra
- AI-Powered System Generates Groundbreaking Research Ideas
Looking ahead, the team envisions AlphaEvolve being used to systematically assess the difficulty of mathematical bounds and conjectures, potentially leading to a new classification system for problems. This work represents a significant step towards automated mathematical discovery, showcasing AlphaEvolve as a powerful new tool for exploring vast search spaces and solving complex optimization problems at scale. For more details, you can read the full research paper here.
