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

Large language models (LLMs) have enabled web agents that follow natural language goals through multi-step browser interactions. However, agents fine-tuned on specific trajectories and domain often struggle to generalize out of domain, and offline training can be compute-inefficient due to noisy, redundant trajectories and long accessibility-tree (AXTree) states. To address both issues, we propose Weasel, a trajectory selection method for offline training of web agents. Weasel selects a fixed-budget subset of trajectory steps by optimizing an objective that balances unary importance with pairwise diversity over states, websites, and interaction patterns, solving efficiently with a greedy algorithm. We further improve efficiency with target-centered AXTree pruning that keeps only content around the ground-truth action target, and we mitigate style mismatch for reasoning-native models by replacing expert traces with model-generated, style-consistent rationales. Across AgentTrek and NNetNav training datasets, evaluations in WebArena, WorkArena, and MiniWob, and experiments with Qwen2.5-7B, Gemma3-4B, and Qwen3-8B, Weasel improves out-of-domain performance while reducing training cost, producing roughly 9.7-12.5$\times$ training speedups over standard fine-tuning. We make the code available at https://github.com/fatemehpesaran310/weasel.

Large language models can now act as web agents that browse websites and complete tasks, such as clicking buttons, filling forms, or searching for information. However, these agents often perform well only on websites similar to those they were trained on. When they face new websites, layouts, or interaction patterns, their performance can drop sharply. Training them is also expensive because web interaction datasets contain many long, noisy, and redundant steps. This paper introduces Weasel, a method that makes web agent training more efficient and more generalizable. Instead of training on all available interaction steps, Weasel selects a small but useful subset of steps that are both important for the task goal and diverse across websites and interaction types. It also shortens web page representations by keeping only the information around the relevant action target. Across multiple web agent benchmarks and model families, Weasel improves performance on unseen domains while reducing training time by roughly 9.7-12.5$\times$ compared to standard full-data fine-tuning. Overall, this work shows that carefully choosing and pruning training data can make web agents more practical, efficient, and robust.