In the current large language model (LLM) field, deep search capabilities have become the "ultimate move" of top intelligent agents. However, this competition has long been dominated by industrial giants with substantial resources. Traditional development models typically rely on resource-intensive pipelines, including pre-training, continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL).

A research team from academia recently released their latest achievement OpenSeeker-v2, completely breaking this conventional understanding. The research report indicates that by using high-quality and high-difficulty task trajectories for training, even a simple supervised fine-tuning (SFT) method can develop a top-performing search agent.

image.png

The team proposed three core optimization strategies in data synthesis: first, expanding the knowledge graph scale to provide a richer exploration space; second, significantly increasing the number of toolkits to extend functional boundaries; finally, implementing strict low-step filtering to ensure the refinement and efficiency of training data.

Experimental data shows that OpenSeeker-v2 (30B scale, ReAct architecture), trained on only 10,600 data points, demonstrated strong dominance in four core benchmark tests: it achieved an accuracy of 46.0% on BrowseComp, 58.1% on BrowseComp-ZH, 34.6% on "Humanity's Last Exam," and as high as 78.0% on xbench. These results not only set new records but also comprehensively surpassed industry models that used complex pipelines of heavy CPT + SFT + RL - such as Tongyi DeepResearch.

image.png

Notably, this is the first state-of-the-art (SOTA) search agent developed by a purely academic team using only SFT technology at the same model scale and architecture. Currently, the team has officially open-sourced the model weights of OpenSeeker-v2. This discovery greatly reduces the R&D threshold for cutting-edge search agents and provides an academic community and open-source community with a more reference-worthy lightweight development path.

Paper URL: https://arxiv.org/pdf/2605.04036