OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories
:::info Stub — Full Engineering Breakdown Coming This paper was featured on Hugging Face Daily Papers on 2026-05-05 with 63 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Yuwen Du et al. |
| Year | 2026 |
| HF Upvotes | 63 |
| arXiv | 2605.04036 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
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.
Engineering Breakdown
Plain English
This paper shows that you can train high-performance search agents using simple supervised fine-tuning (SFT) on carefully synthesized training data, rather than the expensive industrial pipeline of pre-training + CPT + SFT + RL. The key trick is generating 'informative and high-difficulty' trajectories by scaling up knowledge graphs, expanding tool sets, and filtering for efficient solutions—trained on just 10.6K examples, they achieve competitive results with frontier models.
Key Engineering Insight
Data quality and difficulty matter more than pipeline complexity for search agents. Three straightforward modifications—larger knowledge graphs, broader tooling, and strict step-count filtering—can outperform the conventional mega-pipeline approach, suggesting that the bottleneck isn't architectural but data-centric.
Why It Matters for Engineers
If you're building search-based AI systems in production, this means you can skip the massive resource overhead of continual pre-training and RL infrastructure. Instead, focus engineering effort on generating harder, more informative training examples—this is cheaper, more controllable, and potentially faster to iterate on than managing complex multi-stage training pipelines.
Research Context
Previous work required industrial-scale resources (pre-training, RL) to create capable search agents, putting frontier performance out of reach for most teams. This paper challenges that assumption by showing SFT alone works if you invest in trajectory quality rather than pipeline stages, making frontier search capabilities more accessible and democratizing development beyond companies with massive compute budgets.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::