

















Abstract: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 this https URL.
| Comments: | ICML 2026. Code is released at this https URL |
| Subjects: | Machine Learning (cs.LG) |
| Cite as: | arXiv:2605.20291 [cs.LG] |
| (or arXiv:2605.20291v2 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2605.20291 arXiv-issued DOI via DataCite |
From: Fatemeh Pesaran Zadeh [view email]
[v1]
Tue, 19 May 2026 09:19:01 UTC (5,252 KB)
[v2]
Tue, 26 May 2026 02:53:26 UTC (5,252 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理,仅供阅读参考。 原文来自 — 版权归原作者所有。