BrowserForge: Scaling Web Episode via Parallel Browser Sandboxes
BrowserForge reports 203,238 verified web-agent trajectories, each from a distinct open-web site.
The paper describes a parallel browser-sandbox system that sources real reachable websites, proposes tasks from pages, solves them, and filters failed runs. Its released agent acts from screenshots only, though page structure is used while synthesizing data. Fine-tuning a compact multimodal model on the corpus raised Online-Mind2Web success from 25.66% to 33.33%. The authors say broader open-web coverage was a key contributor to the gain. ArXiv · AI/CL/LG's note
The paper describes a parallel browser-sandbox system that sources real reachable websites, proposes tasks from pages, solves them, and filters failed runs. Its released agent acts from screenshots only, though page structure is used while synthesizing data. Fine-tuning a compact multimodal model on the corpus raised Online-Mind2Web success from 25.66% to 33.33%. The authors say broader open-web coverage was a key contributor to the gain. ArXiv · AI/CL/LG's note
score 6