🌐 Explore to Evolve: Scaling Evolved Aggregation Logic via Proactive Online Exploration for Deep Research Agents

🌟 Introduction

• Explore to Evolve aims to generate diverse, high-quality training data for web agent foundation models, enhancing their capabilities in multi-tool usage, information seeking, and information aggregation.

• WebAggregator, the finetuned model on WebAggregatorQA, demonstrates strong performance on GAIA-text and the WebAggregatorQA test set.

---

✨ Features

• 🤖 Fully Automated and Verifiable QA Construction
• 😄 Open Source: Complete codebase including QA construction engine, queries, trajectories, and models.
• 👍 Highly Customizable: Collect data tailored to your needs with minimal human effort, and easily customize your own agent!

---

⚡ Quick Start

Follow these steps to get started:

1️⃣ Clone the Repository

git clone https://github.com/Tencent/WebAggregator

2️⃣ Install Dependencies

1. This project builds upon smolagents’ “open deep research” example 👉 smolagents open_deep_research dependencies. Thanks for their great work and please cite them!

2. Install this project’s requirements:

pip install -r requirements.txt

3. Please note: the implementation must utilize the ./smolagents, which provides the added functionality for trajectory collection by us. Or you can directly replace the smolagets/agents.py in your original library.

---

🚀 Usage

⚙️ Configuration

​Set the configuration in the following files:​​

• ​./config.py: Contains settings for your agent's foundation LLM, the LLMs for specific tools, and dataset paths.
• ./model_list.py: This file is used to implement the method for calling your foundation models (e.g., via vLLM, LiteLLM, or Azure). It calls the models that are configured in ./config.py. We provide an example implementation. For more details, please refer to the smolagents repository.

The function of others:

• ./web_tools.py: Tools for agent. You could modify it to suit your needs.
• ./run_agent.py: The implemented agent.
• ./run: Scripts for running the agent.
• ./data: Input data for QA construction (URLs), evaluation (Benchmarks) and traj sampling (QAs).

---

▶️ Running the Project

Note: Before running any scripts, ensure all paths, model checkpoints, and other necessary parameters are properly set in the source files.

---

1️⃣ Evaluation

To evaluate your agent, serve your tuned checkpoint and update the corresponding settings in config.py. Make sure the correct model_id is set in the evaluation script test.sh, then run:

bash run/test.sh

This command evaluates your specified model and benchmark. After evaluation, it uses LLM-as-judge to assess performance and prints the accuracy.

---

2️⃣ QA Construction

Start building automatic web agent data:

1. Download our collected URLs 👉 URLs or gather URLs related to your domains of interest!

2. Then, run the following command to collect the data.

bash run/QA_building.sh

---

3️⃣ Trajectory Sampling

Training trajectories for fine-tuning your agent foundation models are available at 👉 WebAggregatorQA. Sample data can be found in ./data/train-samples for initial testing purposes.

bash run/traj_sampling.sh

---

Friendly links to other works from Tencent AI Lab

• Deep Research Agent framework: Cognitive Kernel-Pro
• Agent Self-Evolving Research, including WebEvolver, WebCoT, WebVoyager, OpenWebVoyager.

Citation

@misc{wang2025exploreevolvescalingevolved,
      title={Explore to Evolve: Scaling Evolved Aggregation Logic via Proactive Online Exploration for Deep Research Agents}, 
      author={Rui Wang and Ce Zhang and Jun-Yu Ma and Jianshu Zhang and Hongru Wang and Yi Chen and Boyang Xue and Tianqing Fang and Zhisong Zhang and Hongming Zhang and Haitao Mi and Dong Yu and Kam-Fai Wong},
      year={2025},
      eprint={2510.14438},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.14438}, 
}

@misc{fang2025cognitivekernelpro,
      title={Cognitive Kernel-Pro: A Framework for Deep Research Agents and Agent Foundation Models Training}, 
      author={Tianqing Fang and Zhisong Zhang and Xiaoyang Wang and Rui Wang and Can Qin and Yuxuan Wan and Jun-Yu Ma and Ce Zhang and Jiaqi Chen and Xiyun Li and Hongming Zhang and Haitao Mi and Dong Yu},
      year={2025},
      eprint={2508.00414},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2508.00414}, 
}