OpenResearcher: An Innovative Approach to Long-Horizon Research Trajectory Development

The development of deep research agents necessitates the creation of extensive trajectories that combine search, evidence gathering, and complex reasoning. Traditional data collection methods often depend on proprietary web APIs, leading to challenges in scalability, stability, and reproducibility. OpenResearcher addresses these issues by providing a fully reproducible pipeline that separates the initial corpus setup from the multi-turn trajectory synthesis process. This system operates entirely offline, utilizing three fundamental browser functions: search, open, and find, across a vast corpus of 15 million documents. By leveraging the GPT-OSS-120B model, OpenResearcher successfully generates over 97,000 trajectories, including many that involve over 100 tool calls. Fine-tuning a 30B-A3B model on these trajectories resulted in a significant accuracy increase of 34 points on the BrowseComp-Plus benchmark, achieving 54.8% accuracy. The offline nature of the environment allows for thorough analysis, yielding valuable insights into the design of deep research pipelines, including effective data filtering and agent configuration strategies. The OpenResearcher pipeline, along with the synthesized trajectories and model checkpoints, is available at https://github.com/TIGER-AI-Lab/OpenResearcher.