OpenSeeker: 11.7k Synthetic Data Brings Open-Source Search Agent Close to Frontier Systems
Submitted on March 16, 2026, OpenSeeker's fully open model and training data, using only 11.7k synthetic samples and SFT, achieves 29.5% on BrowseComp, higher than open-source DeepDive's 15.3%, and 48.4% on BrowseComp-ZH.
Search agents rely on both model capability and high-quality trajectory data. OpenSeeker uses controllable QA synthesis and trajectory denoising to narrow the gap between open-source and industrial systems, providing a reproducible path for Chinese deep search.
The method performs topological expansion and entity confusion on the Web graph to generate multi-hop questions with controllable coverage and complexity; then uses retrospective summarization to remove redundant and erroneous actions from teacher trajectories. A single training covers BrowseComp, BrowseComp-ZH, xbench-DeepSearch, and WideSearch simultaneously.
Deep search will shift from closed-source product features to an open ecosystem competition involving models, data, and evaluation; training data transparency also makes it easier for enterprises to audit sources, reproduce capabilities, and perform domain adaptation.
When developing search agents, first reuse open data to establish a baseline, then supplement with vertical questions and real evidence, avoiding costly full-trajectory collection from the start.
Need to examine the gap between synthetic tasks and real research, data licensing, search engine dependency, and the safety and factual stability of open models in long tasks.