AI-Scientist-v2: Automating Scientific Discovery

Sakana AI has released AI-Scientist-v2, an autonomous research system that automates the entire scientific process from hypothesis generation to paper writing using large language models and agentic workflows.

Details

AI-Scientist-v2 is a significant advancement in AI-driven research automation. It builds upon the original AI-Scientist by introducing agentic tree search, a method for exploring research directions more effectively than linear approaches. The system maintains a tree of research states, where each node represents a potential research direction. Promising branches are explored deeply while unpromising paths are pruned. This multi-agent architecture enables parallel exploration of research directions, accelerating the discovery process. Sakana AI has evaluated AI-Scientist-v2 across multiple scientific domains, including machine learning, materials science, and computational biology, with impressive results. For example, in machine learning research, the system discovered novel neural architecture components that improved ImageNet accuracy. In materials science, it proposed candidate materials for battery electrolytes with promising computational screening. While AI-Scientist-v2 does not replace human researchers, it augments their capabilities by offering speed, scale, and objectivity. However, the system lacks physical intuition and real-world context, and its creativity is bounded by training data patterns.