Main takeaway: Sakana AI unified over ten black-box optimization methods into one mathematical framework and identified two key design choices: favoring stability vs. peak performance, and converging to one solution vs. exploring multiple. They proposed two hybrid approaches: es-ovi lets developers tune a parameter to prefer stability or peak scores, while adapol and schedpol let algorithms switch dynamically between single-solution convergence and multi-solution exploration.
币界网消息，Sakana AI团队在入选ICML 2026的论文中，首次将十余种黑盒优化算法统一到同一个数学框架下，并指出它们的区别主要来自两个设计选择：一是更追求稳定可靠，还是追求最高性能二是所有搜索都朝同一个答案收敛，还是同时探索多个可能的答案。基于这一发现，研究团队提出了两类混合优化算法，其中es-ovi允许开发者通过一个参数，自由决定算法更偏向「求稳」还是「冲高分」。另一类算法（adapol、schedpol）则允许算法在「所有搜索都朝同一个答案收敛」和「同时探索多个可能答案」之间动态切换。