Evolutionary Intelligence Advances Autonomous Scientific Discovery Systems

The field of artificial intelligence is moving towards autonomous systems for scientific discovery, shifting from task-specific workflows to open-ended exploration guided by experimental and human feedback. Evolutionary computation (EC) offers a strong foundation for such feedback-driven discovery, as its population-based search can maintain diverse candidates and direct exploration using accumulated evidence. However, EC typically focuses on refining candidates for predefined problems, lacking mechanisms for retaining experience over time. To address this limitation, this review proposes Evolutionary Intelligence (EI) for scientific discovery. EI describes AI systems that maintain continuous exploration by linking candidate refinement with the retention of learned experience across multiple evolutionary cycles. The authors introduce a five-dimensional framework to analyze these systems, considering what evolves, how candidates change, why they are selected, where feedback originates, and when evolution occurs. This framework clarifies how EI transforms isolated search trajectories into cumulative scientific insight. The paper illustrates this paradigm across various discovery modes, from evolving specific scientific entities to orchestrating automated research workflows. It also identifies key challenges, including evaluation methods, process traceability, and the need for shared infrastructure, providing a roadmap for advancing from EC to EI in scientific discovery.