Weasel: Out-of-Domain Generalization for Web Agents via Importance-Diversity Data Selection

In three linesWeasel is a trajectory selection method for offline training of web agents. It optimizes a balance between importance and diversity across states, websites, and interaction patterns, with target-centered AXTree pruning. On WebArena, WorkArena, and MiniWob, it improves out-of-domain generalization with 9.7-12.5× training speedups over standard fine-tuning on Qwen2.5-7B, Gemma3-4B, and Qwen3-8B.

Summary generated by Claude — human-verified