AutoSurfer: A New Method for Generating Web Agent Training Data

A team of researchers has unveiled AutoSurfer, an advanced web trajectory generator aimed at tackling the lack of high-quality training data for multimodal large language model (LLM) web agents. Existing methods for automatic trajectory generation often fall short, primarily due to their reliance on homepage-based task proposals or random-walk exploration, leading to vague or fabricated task generation. AutoSurfer addresses these challenges by implementing three main innovations: a systematic breadth-first exploration approach that keeps track of discovered pages and action traces, knowledge propagation to minimize redundant exploration, and the recursive expansion of multi-level graphical user interface elements. This method mimics human browsing patterns, resulting in more accurate and comprehensive trajectory generation. The research paper can be found on arXiv with the identifier 2604.27253.