超代理

Self-improving AI systems aim to reduce reliance on human engineering by learning to enhance their own learning and problem-solving processes. Existing self-improvement approaches depend on fixed, handcrafted meta-level mechanisms, which fundamentally limit the speed of improvement. The Darwin Gödel Machine (DGM) demonstrates open-ended self-improvement in coding by repeatedly generating and evaluating self-modified variants. Since both evaluation and self-modification are coding tasks, improvements in coding ability can lead to enhancements in self-improvement capability. However, this alignment does not generally extend beyond coding domains. Hyperagents are introduced as self-referential agents that integrate a task agent, which is responsible for solving specific tasks, with self-improvement mechanisms, aiming to achieve broader self-improvement capabilities.