Daily AI Paper

Web agents are getting better, but they still need huge amounts of training data, and collecting realistic browser interactions is expensive and narrow. This paper tackles that bottleneck by building a large-scale data factory for web navigation. The core idea is BrowserForge: run hundreds of browser sandboxes in parallel across the open web, use one agent to propose tasks from real pages, another to solve them, and then filter the successful trajectories into training data. Because the system sources pages from hundreds of thousands of publicly reachable websites, the resulting dataset is much broader than prior collections. It ends up with over 203,000 trajectories, each from a distinct website, while the final agent still acts only from screenshots, not from privileged page structure. Why does this matter? Better and more diverse web interaction data translates into stronger multimodal agents, and the paper shows that fine-tuning on BrowserForge improves performance on both live and benchmark web tasks. It is a practical step toward web agents that generalize beyond a handful of familiar sites.