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
arXiv:2605.04036 · 72▲ · cs.AI, cs.CL
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
The paper trains a 30B search agent (OpenSeeker-v2, based on Qwen3-30B-A3B-Thinking) using only supervised fine-tuning on synthetic ReAct trajectories, skipping the usual continual pre-training and reinforcement learning stages. The main contribution is three tweaks to the data synthesis pipeline: expanding the knowledge graph size used to generate questions, enlarging the tool set, and filtering out any trajectory solvable in too few tool-call steps to enforce a difficulty floor.
Why it matters
It suggests an academic team can build a competitive deep-search agent with a small, carefully curated dataset (10.6k trajectories) and one SFT run, avoiding the expensive multi-stage CPT plus SFT plus RL pipeline that industry labs use. For practitioners building agent systems, the takeaway is that trajectory data quality and difficulty may matter more than training-stage complexity.
Practical takeaway
Watch for the open-sourced model weights and data recipe; if you are assembling agent training data, filtering out easy examples and generating tasks that require multi-hop reasoning over a larger graph may give more gains than adding an RL stage. Note the average training trajectory here is 64.67 tool-call steps, so this targets genuinely long-horizon search tasks.
Key result
46.0% on BrowseComp versus 43.4% for Tongyi DeepResearch (trained with full CPT plus SFT plus RL), plus 58.1% BrowseComp-ZH, 34.6% HLE, and 78.0% xbench, all as reported by the authors comparing against baselines taken from other papers' technical reports and leaderboards rather than head-to-head reruns.
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.