Research map / Multi-agent systems and games / LLM agents
WebRetriever: A Large-Scale Comprehensive Benchmark for Efficient Web Agent Evaluation
Wei Dong, Tianyu Fu, Zhe Yu, Hanning Wang, Anyang Su, Zhizhou Fang, Yuyang Chen, Shuo Wang, Minghui Wu, Ping Jiang, Zhen Lei, Chenxu Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add operational documentation (Gemini-2.5-Pro) (ablation) | Protocol I agent · WebRetriever | success rate (%): 40.9 -> 49.2 | Improved |
| add operational documentation (Claude-4.5) (ablation) | Protocol I agent · WebRetriever | success rate (%): 31.3 -> 39.7 | Improved |
| add operational documentation (Gemini-2.5-Pro) (ablation) | Protocol II agent · WebRetriever | success rate (%): 41.4 -> 50.1 | Improved |
| add operational documentation (Claude-4.5) (ablation) | Protocol II agent · WebRetriever | success rate (%): 31.9 -> 40.1 | Improved |
| require information extraction after navigation (Gemini-2.5-Pro) (ablation) | Protocol III agent · WebRetriever | success rate (%): 43.0 -> 21.0 | Hurt |
| require information extraction after navigation (Claude-4.5) (ablation) | Protocol III agent · WebRetriever | success rate (%): 34.0 -> 16.0 | Hurt |
| add rule-based filtering of network requests (ablation) | NavEval · WebRetriever | average human agreement rate (%): 87.8 -> 91.2 | Improved |
| use Claude-4.5-Sonnet instead of GPT-4o as judge (ablation) | NavEval · WebRetriever | average human agreement rate (%): 86.4 -> 91.5 | Improved |
| use Claude-4.5-Sonnet instead of O4-mini as judge (ablation) | NavEval · WebRetriever | average human agreement rate (%): 89.0 -> 91.5 | Improved |
| add network request signals (NavEval vs WebJudge) | LLM-as-judge · WebRetriever | average human agreement rate (%): 81.0 -> 91.2 | Improved |
| NavEval vs best baseline judge on Online-Mind2Web | LLM-as-judge · Online-Mind2Web | average human agreement rate (%): 87.2 -> 97.0 | Improved |
| move from navigation-only (Protocol I) to end-to-end task completion (Protocol III) | web agents · WebRetriever | average success rate (%): 21.1 -> 11.8 | Hurt |
Closest papers by meaning
- REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.93)
- TimeWarp: Evaluating Web Agents by Revisiting the Past (NeurIPS 2026 · similarity 0.90)
- AgentWebBench: Benchmarking Multi-Agent Coordination in Agentic Web (ICML 2026 · similarity 0.90)
- ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents (NeurIPS 2026 · similarity 0.89)
- Go-Browse: Training Web Agents with Structured Exploration (ICLR 2026 · similarity 0.89)
- Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
