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

2026-05-22 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, long

Summary

Weasel is a novel trajectory selection method designed to enhance out-of-domain generalization and computational efficiency for large language model (LLM)-based web agents. Addressing issues like poor generalization from fine-tuned agents and compute-intensive offline training due to noisy, redundant data, Weasel selects a fixed-budget subset of trajectory steps. It optimizes an objective balancing unary importance with pairwise diversity across states, websites, and interaction patterns, solved efficiently via a greedy algorithm. The method also incorporates target-centered AXTree pruning, which retains only action-relevant content, and replaces expert traces with model-generated, style-consistent rationales for reasoning-native models. Evaluated across AgentTrek and NNetNav datasets, and benchmarks like WebArena, WorkArena, and MiniWob with models such as Qwen2.5-7B, Gemma3-4B, and Qwen3-8B, Weasel demonstrates 9.7-12.5x training speedups and improved out-of-domain performance, including a +4.8 gain in zero-shot transfer for Qwen3-8B.

Key takeaway

For Machine Learning Engineers developing web agents that must generalize across diverse online environments, you should re-evaluate traditional offline training methods. Weasel demonstrates that curating training data for importance and diversity, combined with target-centered AXTree pruning and self-reasoning synthesis, can yield 9.7-12.5x training speedups and substantially improve out-of-domain performance. Integrate these data selection and processing techniques to build more robust and efficient web agents.

Key insights

Weasel enhances web agent generalization and efficiency through importance-diversity trajectory selection and state pruning.

Principles

• Prioritize data diversity to mitigate overfitting.
• Balance step importance with state diversity for curation.
• Ensure reasoning trace style matches the fine-tuned model.

Method

Weasel employs a greedy algorithm for fixed-budget trajectory subset selection, optimizing unary importance and pairwise diversity, complemented by AXTree pruning and self-reasoning synthesis.

In practice

• Implement importance-diversity data selection for web agents.
• Use target-centered AXTree pruning for state efficiency.
• Synthesize style-consistent reasoning traces for LLM fine-tuning.

Topics

• Web Agents
• Out-of-Domain Generalization
• Trajectory Selection
• Data Curation
• AXTree Pruning
• Large Language Models

Code references

• fatemehpesaran310/weasel

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• ARMOR-MAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning — Adaptive control and genuine model heterogeneity are crucial for efficient and accurate multi-agent LLM reasoning.
• OpenThoughts-Agent: Data Recipes for Agentic Models — OpenThoughts-Agent provides an open data pipeline and insights for training agentic models that generalize across diverse tasks.
• Intelligent sampling in Microsoft Foundry: the science behind selecting better production traces — Diversity sampling using MinHash farthest-first traversal significantly enhances dataset lexical variety and LLM judge preference for agent evaluation and fine-tuning.
• AgentKernelArena: Generalization-Aware Benchmarking of GPU Kernel Optimization Agents — AgentKernelArena benchmarks AI agents on GPU kernel optimization, revealing strong performance but critical generalization gaps on unseen input configurations.
• Scaling Web Agent Training through Automatic Data Generation and Fine-grained Evaluation — A constraint-based evaluation framework enables scalable, high-quality web agent training by leveraging partially successful trajectories.
• Co-Scraper: query-aware DOM Pruning and Reusable Scraper Synthesis for Lightweight Web Data Extraction — Query-aware DOM pruning and stable strategy induction enable reusable web scrapers for scalable data extraction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.