Google DeepMind researcher presents WebAgent for digital automation

İzzeddin Gür, a Research Scientist at Google DeepMind, will present research on WebAgent at Bilkent University. The presentation focuses on the development of autonomous digital agents capable of operating in real-world environments.

WebAgent aims to bridge the gap between Reinforcement Learning (RL) and Large Language Models (LLMs). While RL provides a foundation for sequential decision-making, it often struggles with the unstructured and noisy state spaces of modern web pages. Conversely, LLMs excel at reasoning but struggle with long-horizon execution. WebAgent addresses these challenges by decomposing tasks into high-level planning, targeted sub-state selection to manage complex HTML code, and program synthesis to enable open-ended actions.