# WebExplorer: Explore and Evolve for Training Long-Horizon Web Agents

**Authors:** Junteng Liu, Yunji Li, Chi Zhang, Jingyang Li, Aili Chen, Ke Ji, Weiyu Cheng, Zijia Wu, Chengyu Du, Qidi Xu, Jiayuan Song, Zhengmao Zhu, Wenhu Chen, Pengyu Zhao, Junxian He  
**arXiv:** [2509.06501](https://arxiv.org/abs/2509.06501)  
**Date:** 2025-09-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2509.06501v1/

## Key points

- WebExplorer is a systematic data generation approach that utilizes model-based exploration and iterative query evolution to address limited data challenges in information seeking.
- The method creates challenging query-answer pairs requiring multi-step reasoning and complex web navigation.
- WebExplorer-8B, developed through supervised fine-tuning and reinforcement learning on a high-quality dataset, supports a context length of 128K and up to 100 tool calling turns for long-horizon problem solving.
- Achieves state-of-the-art performance across various information-seeking benchmarks, outperforming larger models on tasks like BrowseComp-en/zh and WebWalkerQA.
- Demonstrates strong generalization on the HLE benchmark despite being trained on knowledge-intensive QA data.
- Proposes a different model-based exploration approach with WebExplorer, leveraging powerful LLMs for autonomous construction of the information space through prompting.
- Showcases exceptional performance in information-seeking tasks and sets a new standard for long-horizon web agents through innovative data generation and modeling techniques.

## Summary

, , , , 

The authors present WebExplorer, a systematic data generation approach that utilizes model-based exploration and iterative query evolution to address the challenge of limited data for information seeking. This method creates challenging query-answer pairs that require multi-step reasoning and complex web navigation. Through leveraging a high-quality dataset, the authors develop the advanced web agent WebExplorer-8B via supervised fine-tuning and reinforcement learning. The model supports a context length of 128K and up to 100 tool calling turns, enabling long-horizon problem solving. WebExplorer-8B achieves state-of-the-art performance across various information-seeking benchmarks, outperforming models of larger sizes on tasks such as BrowseComp-en/zh and WebWalkerQA. Additionally, the model demonstrates strong generalization on the HLE benchmark despite being trained on knowledge-intensive QA data. The results highlight WebExplorer-8B as a practical approach towards developing long-horizon web agents. Furthermore, the authors propose a different model-based exploration approach with WebExplorer, leveraging powerful LLMs to autonomously construct the information space through prompting. By providing a seed entity and example QA pairs, the model conducts iterative search and browsing actions to synthesize challenging query-answer pairs that require deep reasoning across multiple connections. Overall, WebExplorer showcases exceptional performance in information-seeking tasks and sets a new standard for long-horizon web agents through its innovative data generation approach and advanced modeling techniques.