OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis
Zhuofeng Li, Dongfu Jiang, Xueguang Ma, Haoxiang Zhang, Ping Nie, Yuyu Zhang, Kai Zou, Jianwen Xie, Yu Zhang, Wenhu Chen
2026-03-24
Summary
This paper introduces OpenResearcher, a new way to create training data for AI agents that need to do complex research online, like searching the web, reading articles, and putting information together to answer questions.
What's the problem?
Training these 'research agents' usually requires a lot of data showing them how to perform these tasks, but getting this data is difficult and expensive. Current methods rely on using real website search engines and APIs, which can be unreliable, change without notice, and aren't easily shared with other researchers to reproduce results.
What's the solution?
The researchers built a system that creates this training data completely offline. They started with a large collection of documents (15 million!), and then used a powerful AI model (GPT-OSS-120B) to simulate a user performing research. This simulation uses only three basic actions – searching, opening a webpage, and finding information on a page – to create over 97,000 complete research 'trajectories'. They then used this data to improve a smaller AI model (30B-A3B).
Why it matters?
This work is important because it provides a reliable and reproducible way to train research agents. By creating the data offline, they avoid the problems of using real-world web APIs. The improved AI model performs significantly better on research tasks, and the system allows researchers to study *how* these agents learn and what makes them successful, leading to better designs in the future. They've also made all their tools and data publicly available so others can build on their work.
Abstract
Training deep research agents requires long-horizon trajectories that interleave search, evidence aggregation, and multi-step reasoning. However, existing data collection pipelines typically rely on proprietary web APIs, making large-scale trajectory synthesis costly, unstable, and difficult to reproduce. We present OpenResearcher, a reproducible pipeline that decouples one-time corpus bootstrapping from multi-turn trajectory synthesis and executes the search-and-browse loop entirely offline using three explicit browser primitives: search, open, and find, over a 15M-document corpus. Using GPT-OSS-120B as the teacher model, we synthesize over 97K trajectories, including a substantial long-horizon tail with 100+ tool calls. Supervised fine-tuning a 30B-A3B backbone on these trajectories achieves 54.8\% accuracy on BrowseComp-Plus, a +34.0 point improvement over the base model, while remaining competitive on BrowseComp, GAIA, and xbench-DeepSearch. Because the environment is offline and fully instrumented, it also enables controlled analysis, where our study reveals practical insights into deep research pipeline design, including data filtering strategies, agent configuration choices, and how retrieval success relates to final answer accuracy. We release the pipeline, synthesized trajectories, model checkpoints, and the offline search environment at https://github.com/TIGER-AI-Lab/OpenResearcher.