EAGER: Beyond Scaling: Scientific Discovery as a Driver of Next-Generation AI

Recent advances in artificial intelligence have been driven by training increasingly large computer models on enormous amounts of data using ever more computing power. This approach has transformed language processing, image analysis, and many other applications and is beginning to accelerate scientific research. However, many scientific discoveries require capabilities that today’s artificial intelligence systems still lack, including the ability to reason about situations that have never been observed and quantify uncertainty. This project will investigate a fundamental question in artificial intelligence: whether continuing to build larger models is sufficient to achieve these capabilities or whether fundamentally new ideas are required. By treating scientific discovery as a laboratory for understanding artificial intelligence itself, the project seeks to identify the foundations of the next generation of artificial intelligence systems. The results will strengthen the nation’s leadership in artificial intelligence and scientific innovation, contribute to more reliable artificial intelligence for science and other high-consequence applications, and help prepare students for interdisciplinary research at the intersection of artificial intelligence and the sciences. The project addresses three research questions. First, it will determine which abilities required for scientific discovery emerge naturally as artificial intelligence models are trained with larger amounts of data and computation, and which remain beyond the reach of scaling alone. Second, it will investigate what additional ingredients are needed to achieve these abilities, including scientific knowledge, structured reasoning, uncertainty awareness, and iterative hypothesis generation and testing. Third, it will develop an evaluation framework for assessing scientific reasoning and the ability to generalize beyond previously observed conditions. The research will develop a scientific-discovery testbed that combines configurable graph-based dynamical systems with scientific process simulators from hydrology, aquatic sciences, and agriculture to create controlled experimental environments for systematically evaluating artificial intelligence systems under varying levels of hidden-state complexity, scientific structure, and distribution shift. The resulting benchmarks, evaluation methodologies, and scientific insights will provide a foundation for future research on next-generation artificial intelligence systems capable of more reliable scientific reasoning and discovery. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.