Source-linked AI summary
OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories
Yuwen Du, Rui Ye, Shuo Tang, Keduan Huang, Xinyu Zhu, Yuzhu Cai, Siheng Chen
TL;DR
Frontier search-agent development is often tied to resource-intensive industrial training pipelines. OpenSeeker-v2 shows that single SFT on difficult, information-rich trajectories can achieve state-of-the-art performance across four benchmarks and outperform heavier-pipeline agents.
Problem
Frontier search agents are typically developed through resource-intensive, multi-stage pipelines dominated by well-funded industrial entities.
Method
OpenSeeker-v2 applies single SFT to high-difficulty, information-rich trajectories synthesized with larger graphs, expanded tool sets, and strict low-step filtering.
Results
OpenSeeker-v2 achieves state-of-the-art performance across four representative benchmarks, including 46.0% on BrowseComp, and outperforms agents trained with heavier pipelines.
Takeaways & Limitations
The findings suggest that simple SFT can rival resource-intensive industrial pipelines when powered by high-quality, difficult, and rich training data.
Takeaways & Limitations
The report limits its scaling claim to internal observations suggesting strong scaling potential for high-quality synthesized data.
Abstract
from arXiv · showhide
Deep search capabilities have become an indispensable competency for frontier Large Language Model (LLM) agents, yet their development remains dominated by industrial giants. The typical industry recipe involves a highly resource-intensive pipeline spanning pre-training, continual pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL). In this report, we show that when fueled with informative and high-difficulty trajectories, a simple SFT approach could be surprisingly powerful for training frontier search agents. By introducing three simple data synthesis modifications: scaling knowledge graph size for richer exploration, expanding the tool set size for broader functionality, and strict low-step filtering, we establish a stronger baseline. Trained on merely 10.6k data points, our OpenSeeker-v2 achieves state-of-the-art performance across 4 benchmarks (30B-sized agents with ReAct paradigm): 46.0% on BrowseComp, 58.1% on BrowseComp-ZH, 34.6% on Humanity's Last Exam, and 78.0% on xbench, surpassing even Tongyi DeepResearch trained with heavy CPT+SFT+RL pipeline, which achieves 43.4%, 46.7%, 32.9%, and 75.0%, respectively. Notably, OpenSeeker-v2 represents the first state-of-the-art search agent within its model scale and paradigm to be developed by a purely academic team using only SFT. We are excited to open-source the OpenSeeker-v2 model weights and share our simple yet effective findings to make frontier search agent research more accessible to the community.
1 Introduction
OpenSeeker-v2 shows that a straightforward SFT approach, trained on informative and high-difficulty trajectories, can achieve frontier search-agent performance without the typical resource-intensive industrial pipeline. Using a 30B ReAct model and only about 10k trajectories, it reaches state-of-the-art results across four benchmarks and is fully open-sourced by an academic team.
- Introduction: OpenSeeker-v2 uses straightforward SFT with high-quality, high-difficulty, and rich training trajectories instead of relying on complex multi-stage training pipelines.The approach shifts emphasis from CPT, SFT, and RL combinations toward the quality of synthesized trajectories.
- Introduction: Scaling graph size, expanding tool-set size, and applying strict low-step filtering make the synthesized trajectories richer, broader, and more difficult.These modifications improve exploration diversity, tool functionality, and trajectory quality.
- Introduction: 46.0% on BrowseComp, 58.1% on BrowseComp-ZH, 34.6% on Humanity’s Last Exam, and 78.0% on xbench establish OpenSeeker-v2 as a new state-of-the-art system across four benchmarks.The results come from training a 30B parameter model with a single SFT run on merely 10k high-difficulty trajectories.
- Introduction: OpenSeeker-v2 is the first state-of-the-art search agent at its model scale and ReAct paradigm developed entirely by a purely academic team using only SFT.The authors also fully open-source its model weights to support reproducible and more accessible frontier search-agent research.
2 Methodology and Results
OpenSeeker-v2 shows that simple SFT can produce strong long-horizon search behavior when trained on informative, difficult trajectories. Its data synthesis scales evidence exploration and tool diversity while filtering out shallow instances, yielding top performance among comparable-scale ReAct agents.
- Data synthesis: Scaling K expands evidence subgraphs, increasing reasoning-path diversity and making synthetic questions more likely to require multi-node evidence aggregation.The expanded graph increases contextual richness and multi-hop dependency in the synthesized data.
- Data synthesis: Expanding A encourages diverse interaction patterns and complementary-tool use, producing more flexible and functionally rich ReAct problem-solving behavior.Each trajectory contains tool-call steps, reasoning traces, observations, and a final answer.
- Data synthesis: Trajectories with T(τ) < Tmin are discarded because they can often be solved by direct lookup or shallow keyword matching.This strict minimum tool-call threshold enforces a difficulty floor before standard SFT.
- Results: 46.0% on BrowseComp, 58.1% on BrowseComp-ZH, 34.6% on Humanity’s Last Exam, and 78.0% on xbench establish OpenSeeker-v2’s strongest overall performance among ∼30B ReAct-based search agents using SFT only.The reported system is OpenSeeker-v2-30B-SFT, evaluated against comparable-scale agents trained with heavier pipelines.
- Results: 64.67 steps per trajectory is OpenSeeker-v2’s average, versus 46.97 for OpenSeeker-v1 and 36.01 for RedSearcher.The longer trajectories indicate more complex multi-step reasoning and longer-horizon information seeking in the training data.
3 Conclusion
The conclusion presents OpenSeeker-v2 as evidence that high-quality, high-difficulty, and rich data can make simple SFT competitive with resource-intensive search-agent training. It highlights three data-collection modifications and identifies scaling synthesized data as the next direction.
- High-quality, high-difficulty, and rich data enables a search agent trained with simple SFT to rival agents trained with extensive resources.
- The data pipeline uses three modifications: scaling graph size, expanding the tool set, and low-step filtering.
- 10.6k samples suffice to train the final search agent, OpenSeeker-v2.
- What’s next: High-quality synthesized data shows strong scaling potential.
- What’s next: Future work will scale data quantity, quality, and diversity to further push search-agent limits.