WebRetriever: A Large-Scale Comprehensive Benchmark for Efficient Web Agent Evaluation

· Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

WebRetriever is a new large-scale benchmark designed to address limitations in existing web agent evaluation frameworks, which suffer from insufficient scale, limited domain diversity, and inadequate assessment of real-world deployment needs. It comprises 1,550 tasks across 800 websites, spanning consumer, professional, and enterprise sectors with comprehensive user intent patterns. Complementing this, NavEval is a novel LLM-as-Judge framework that utilizes rich interaction context beyond visual screenshots, achieving over 90% alignment with human judgment. The benchmark introduces three evaluation protocols: navigation proficiency, knowledge-assisted interaction, and end-to-end task completion with information extraction. Experimental analysis reveals significant performance disparities, with agents achieving only 21.1% success rate for basic navigation (Protocol I) and 11.8% for end-to-end tasks (Protocol III). Operational documentation improved success rates to 29.2% on average. WebRetriever provides fine-grained diagnostic insights, highlighting current agents' inability to reliably handle practical end-to-end web tasks.

Key takeaway

For AI Engineers evaluating web agents for real-world deployment, you must move beyond basic navigation success metrics. Your evaluation framework should incorporate diverse, large-scale benchmarks like WebRetriever, covering varied domains and user intents. Crucially, assess end-to-end task completion, including information extraction, and integrate operational knowledge to improve agent reliability. Rely on robust automated evaluation methods like NavEval, achieving over 90% human agreement, for fine-grained diagnostic insights.

Key insights

Web agent evaluation requires large-scale, diverse benchmarks and fine-grained, multi-modal LLM-as-Judge methods to reflect real-world performance.

Principles

Method

NavEval integrates task descriptions, URLs, filtered web requests, executed actions, and final screenshots, applying rule-based filtering to enhance LLM-as-Judge accuracy.

In practice

Topics

Code references

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

Related on AIssential

Open in AIssential →

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