Explorer: Scaling Exploration-driven Web Trajectory Synthesis for Multimodal Web Agents
Vardaan Pahuja, Yadong Lu, Corby Rosset, Boyu Gou, Arindam Mitra, Spencer Whitehead, Yu Su, Ahmed Hassan Awadallah
- 🏛 Institutions
- MSR, Redmond, OSU
- 📅 Date
- July 2025
- 📑 Publisher
- Findings of ACL 2025
- 💻 Env
- Web
- 🔑 Keywords
TLDR
This paper targets the shortage of large, diverse web-agent trajectories by synthesizing a 94K-trajectory multimodal dataset through scalable exploration and refinement over 49K URLs. Training on this dataset yields a strong web agent, Explorer, and shows that data scaling is a major driver of web-agent performance.
Related papers
- OS-Genesis: Automating GUI Agent Trajectory Construction via Reverse Task SynthesisDecember 27, 2024 · ACL 2025
- On the Effects of Data Scale on UI Control AgentsJune 6, 2024 · NeurIPS 2024 Datasets and Benchmarks Track
- WebForge: Breaking the Realism-Reproducibility-Scalability Trilemma in Browser Agent BenchmarkApril 13, 2026 · arXiv
- MolmoWeb: Open Visual Web Agent and Open Data for the Open WebApril 9, 2026 · arXiv
- WebArena-Infinity: Generating Browser Environments with Verifiable Tasks at ScaleMarch 2026 · Blog Post
- HATS: Hardness-Aware Trajectory Synthesis for GUI AgentsMarch 12, 2026 · CVPR 2026