AI与数学“双向奔赴”，中国团队突破亲吻数问题

Core Viewpoint - A breakthrough in the kissing number problem has been achieved by a collaborative research team from the Shanghai Institute of Science and Intelligence, Peking University, and Fudan University, utilizing the PackingStar reinforcement learning system to explore complex high-dimensional spaces and surpass previous human-known optimal structures [1][4]. Group 1: Research Breakthroughs - The research team has broken the known optimal kissing number structures in dimensions 25-31, as well as the long-standing values for 14 and 17 dimensions for "two-ball kissing numbers," and 12, 20, and 21 dimensions for "three-ball kissing numbers" [1]. - The PackingStar system has transformed the complex high-dimensional geometric problem into an algebraic problem compatible with GPU parallel logic, significantly enhancing the computational potential of AI models [4][6]. Group 2: AI and Mathematics Integration - The research represents a bidirectional attempt between AI and mathematics, with AI advancements lowering the barriers for researchers to tackle complex problems [2]. - Previous attempts to use AI in the kissing number problem, such as DeepMind's AlphaEvolve, resulted in limited breakthroughs, highlighting the unique approach of the PackingStar system in achieving substantial improvements [3][4]. Group 3: Future Plans and Applications - The research team plans to improve the system further, expand to the entire space of sphere packing problems, explore applications in graph theory, and collaborate with more mathematicians [8]. - The results from the PackingStar project have led to a 2-3 times increase in search efficiency and saved over 100,000 GPU hours, establishing a reusable interdisciplinary intelligent computing paradigm [6].