DRIVE: A Dual-Level Skill Model for Web Agents Under Continual Learning

A recent study, arXiv:2605.23939, presents DRIVE, a dual-level skill framework tailored for web agents engaged in continual learning. This model tackles the essential distinction between high-level reasoning knowledge, such as task decomposition (e.g., finding routes for flight bookings), and low-level interaction knowledge, like clicking specific buttons on websites. Current approaches tend to store experiences uniformly, resulting in a conflict: abstract representations struggle with executable tasks on specific pages, while concrete representations do not generalize well across different domains. By separating these two knowledge types, DRIVE enhances agents' ability to develop skills more efficiently on new websites. The research can be accessed on arXiv.