Please note: This PhD seminar will take place online.
Dongfu Jiang, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Wenhu Chen
Deep research agents aim to answer complex information-seeking questions by iteratively searching for evidence, reading documents, aggregating information, and reasoning over many steps. However, existing training pipelines often depend on proprietary web APIs, making large-scale data synthesis expensive, unstable, and difficult to reproduce. In this talk, I will present OpenResearcher, a fully open and reproducible pipeline for synthesizing long-horizon deep research trajectories. OpenResearcher decouples one-time corpus construction from multi-turn trajectory synthesis, and runs the search-and-browse loop entirely offline over a 15M-document corpus using three simple browser primitives: search, open, and find. Using a large teacher model, we synthesize over 97K research trajectories, including many long-horizon examples with 100+ tool calls, and use them to train a 30B-A3B agent model. I will discuss the design of the offline environment, trajectory synthesis and filtering, model training, and empirical findings on deep research benchmarks. I will also highlight practical lessons about tool-space design, data quality, and the relationship between retrieval success and final answer accuracy.