Toward Autonomous Digital Agents in the Real World
İzzeddin Gür
Google Deepmind
Abstract: Reinforcement Learning (RL) provides a mathematically grounded foundation for sequential decision-making. Yet, deploying RL agents for real-world digital tasks remains a significant challenge. Contemporary web pages present unstructured state spaces that exhibit long-context HTML code with high perceptual noise, lack of reward feedback, and active safety measures. On the other hand, Large Language Models (LLMs) excel at few-shot reasoning and local task composition but struggle with long-horizon execution and unaligned multi-step instructions. In this talk, I present WebAgent, a digital agent that combines foundational RL decision making with LLM program synthesis for end-to-end automation. WebAgent addresses compositional complexity of real-world digital tasks by breaking decision-making into three components: (1) high-level task decomposition and planning, (2) targeted sub-state selection to simplify long HTML code, and (3) program synthesis to replace fixed action spaces with open-ended programs via generalist LLMs. Empirical results demonstrate that explicitly structuring task decomposition and planning enables scalable generalization across real-world web environments.
Biography: İzzeddin Gür is a Research Scientist at Google Deepmind, where his research focuses on reinforcement learning, autonomous agents, and decision-making algorithms. His work investigates foundational and practical aspects of agentic systems, including latent-space dynamics, post-training, and web agent architectures. İzzeddin Gür’s research has been published in premier artificial intelligence and machine learning conferences such as ICLR, NeurIPS, ICML and highlighted in popular science outlets such as Science magazine. Beyond research, he actively serves as an Area Chair for NeurIPS and ACL.
DATE: October 22, 2026 Thursday @ 10:30
Place: EA 502