Recent progress in large language models, reinforcement learning, and agentic search has sparked growing interest in automating aspects of AI research. While current AI research agents can reproduce results, perform hill-climbing optimization, and iteratively improve models, they still lack the ability to generate truly novel, meaningful, and high-impact scientific insights. This talk examines the core ingredients required for autonomous AI research—data, environments, training algorithms, and evaluation—and argues that despite encouraging progress, current systems are missing the discovery process that drives genuine innovation. We position open-ended learning, exploration-driven search, and better evaluation frameworks as key steps toward building AI scientists capable of sustained research progress and true scientific discovery.
Roberta Raileanu is a Senior Staff Research Scientist at Google DeepMind and Honorary Lecturer at UCL. Her work focuses on designing open-ended learning systems drawing from different fields such as reinforcement learning, self-supervised learning, evolutionary search, and foundation models. Previously, Roberta was a Research Scientist at Meta, where she led tool use for Llama models and worked on applications to scientific discovery and accelerating AI research itself. Roberta received her PhD in computer science at NYU, where she worked on generalization in deep reinforcement learning.