陶哲轩力推AlphaEvolve：解决67个不同数学问题，多个难题中超越人类最优解

Core Viewpoint - AlphaEvolve is presented as a powerful new tool for mathematical discovery, capable of autonomously discovering novel mathematical constructs and surpassing existing human optimal results in certain problems [2][5]. Group 1: AlphaEvolve's Capabilities - AlphaEvolve has been tested on 67 mathematical problems across various fields, including combinatorial mathematics, geometry, mathematical analysis, and number theory [4]. - The system not only reproduces many known optimal solutions but also demonstrates unique discovery capabilities, including the ability to autonomously find new mathematical constructs previously unseen by humans [6][7]. - In the Nikodym set problem, AlphaEvolve provided a preliminary construct that, while not optimal, served as an excellent intuitive jumping-off point for human researchers, leading to an improved known upper bound [8]. Group 2: Performance Metrics - AlphaEvolve outperforms traditional tools in scalability, robustness, and interpretability [9]. - In the arithmetic Kakeya conjecture, the system improved a known lower bound from 1.61226 to 1.668 and inspired mathematicians to establish new asymptotic relationships [12]. - The system's ability to generate clear and interpretable program code allows human experts to analyze and extract general mathematical formulas from its findings [12]. Group 3: Problem-Solving Techniques - AlphaEvolve effectively handles high-dimensional parameter spaces, complex geometric constraints, and Monte Carlo simulation-based scoring functions [20][21]. - In a minimum triangle density problem, the system utilized the non-convexity of the problem space to achieve scores beyond theoretical optimality, prompting researchers to design a more robust scoring function [24]. - The system demonstrated excellent generalization capabilities by discovering a universal construct that achieves optimal results for all perfect square inputs [29]. Group 4: Operational Modes - AlphaEvolve operates in two main modes: "search mode" for efficiently discovering optimal mathematical constructs and "generalizer mode" for creating universal programs applicable to any given parameter [32][33]. - In search mode, the system evolves heuristic algorithms that can trigger large-scale, inexpensive computations to explore millions of candidate constructs [35]. - The generalizer mode challenges the system to identify patterns from optimal solutions found at small scales and generalize them into a universal formula or algorithm [37]. Group 5: Human-AI Collaboration - The efficiency of AlphaEvolve is significantly enhanced by expert guidance, indicating a high sensitivity to human input [31]. - The system's architecture supports parallelization, allowing researchers to explore multiple problem instances simultaneously, which is particularly effective for multi-parameter geometric problems [31].