DeepMind Explores AI Agents for Scientific Discovery

Roberta Rail from Google DeepMind presented insights at raais2026 regarding the role of AI agents in scientific discovery. She highlighted that while AI agents possess self-improvement mechanisms, they often reach a performance plateau relatively quickly. This contrasts with human researchers, who consistently demonstrate the ability to make significant conceptual leaps and breakthroughs. To address this disparity, Rail's work explores several innovative approaches. These include applying reinforcement learning techniques to uncover scientific "Move 37" moments—referencing a famous Go move by AlphaGo—and employing evolutionary search methods that specifically reward novelty in findings. Furthermore, she introduced DiscoBench, a comprehensive benchmark comprising over 400 million research tasks, designed to rigorously test and advance AI's capabilities in discovery. Her presentation underscored the critical difference between incremental scaling and fundamental conceptual innovation, drawing a parallel to how an abacus, despite its utility, never evolved into a computer.