Evolutionary Intelligence: New Framework for Autonomous Scientific Discovery

In a recent arXiv preprint (2607.09025), researchers Chao Wang, Lingling Li, Fang Liu, and Licheng Jiao propose a new paradigm called evolutionary intelligence (EI) for scientific discovery. They argue that while AI is shifting from task-specific workflows toward autonomous systems, current evolutionary computation (EC) methods are limited because they primarily refine candidates for predefined problems without retaining experience. EI bridges this gap by combining candidate refinement with experience retention across multiple evolutionary cycles, enabling cumulative discovery.

The paper introduces a five-dimensional analytical framework that asks what evolves, how candidates change, why they are selected, where feedback originates, and when evolution occurs. This framework helps transform isolated search trajectories into cumulative scientific insight. The authors demonstrate EI across diverse discovery modes—from evolving concrete scientific entities to orchestrating automated research workflows. They also identify critical bottlenecks in evaluation, process traceability, and shared infrastructure, providing a roadmap for moving from EC to EI. This work is a perspective article submitted to a Springer Nature journal.