Beyond Pixels: Exploring DOM Downsampling for LLM-Based Web Agents
The paper 'Beyond Pixels' addresses the limitation of using raw, large HTML (DOM) for LLM-based web agents. It introduces D2Snap, a method of intelligently downsampling the DOM by merging elements and summarizing text. This approach achieves a 73% success rate on web tasks, significantly outperforming traditional screenshot methods while drastically reducing token usage.