DRIVE: Modeling Skills at the Reasoning and Interaction Levels for Web Agents under Continual Learning
Researchers have developed a new framework called DRIVE to improve the performance of web agents. DRIVE disentangles reasoning skills, which are abstract and transferable, from interaction skills, which are page-specific and executable. This separation allows agents to better learn and adapt to new websites by recognizing reusable task logic while grounding actions to specific page elements. Experiments show DRIVE significantly outperforms skill-free baselines in task success rates across multiple domains. AI
IMPACT Enhances web agent capabilities by improving task decomposition and page element manipulation, potentially leading to more sophisticated automated web interactions.