OpenResearcher: a reproducible and scalable pipeline for training deep research agents

Introduces OpenResearcher, a pipeline for training deep research agents. It addresses the data bottleneck by building a reproducible offline research environment over 15 million documents, using GPT-OSS-120B to generate over 97,000 long-horizon research trajectories, and distilling these into a smaller Nemotron-3-Nano-30B-A3B model. The pipeline involves building the environment, generating trajectories with tools like search, open, and find, and then distilling the knowledge into a smaller model. Demonstrates significant performance gains on benchmarks like BrowseComp-Plus, outperforming larger proprietary models.