Source-linked AI summary
OpenSeeker: Democratizing Frontier Search Agents by Fully Open-Sourcing Training Data
Yuwen Du, Rui Ye, Shuo Tang, Xinyu Zhu, Yijun Lu, Yuzhu Cai, Siheng Chen
TL;DR
High-performance search-agent research is limited by scarce transparent, high-quality training data. OpenSeeker addresses this with controllable fact-grounded QA synthesis and denoised trajectory synthesis, achieving state-of-the-art results across multiple benchmarks with 11.7k samples and releasing its data and weights openly.
Problem
High-performance search-agent development remains dominated by industry because transparent, high-quality training data is scarce.
Method
OpenSeeker combines fact-grounded scalable controllable QA synthesis with denoised trajectory synthesis and releases the resulting training data and model weights.
Results
OpenSeeker achieves state-of-the-art performance across frontier search benchmarks using 11.7k synthesized samples and a single SFT run.
Takeaways & Limitations
Full release of the synthesis pipeline, high-fidelity training data, and model weights provides resources for the community to replicate and build upon frontier-level search capabilities.
Takeaways & Limitations
Resource constraints limited training to a single run, restricting verification on more challenging data and exploration of parameters and filtering strategies.
Abstract
from arXiv · showhide
Deep search capabilities have become an indispensable competency for frontier Large Language Model (LLM) agents, yet the development of high-performance search agents remains dominated by industrial giants due to a lack of transparent, high-quality training data. This persistent data scarcity has fundamentally hindered the progress of the broader research community in developing and innovating within this domain. To bridge this gap, we introduce OpenSeeker, the first fully open-source search agent (i.e., model and data) that achieves frontier-level performance through two core technical innovations: (1) Fact-grounded scalable controllable QA synthesis, which reverse-engineers the web graph via topological expansion and entity obfuscation to generate complex, multi-hop reasoning tasks with controllable coverage and complexity. (2) Denoised trajectory synthesis, which employs a retrospective summarization mechanism to denoise the trajectory, therefore promoting the teacher LLMs to generate high-quality actions. Experimental results demonstrate that OpenSeeker, trained (a single training run) on only 11.7k synthesized samples, achieves state-of-the-art performance across multiple benchmarks including BrowseComp, BrowseComp-ZH, xbench-DeepSearch, and WideSearch. Notably, trained with simple SFT, OpenSeeker significantly outperforms the second-best fully open-source agent DeepDive (e.g., 29.5% v.s. 15.3% on BrowseComp), and even surpasses industrial competitors such as Tongyi DeepResearch (trained via extensive continual pre-training, SFT, and RL) on BrowseComp-ZH (48.4% v.s. 46.7%). We fully open-source the complete training dataset and the model weights to democratize frontier search agent research and foster a more transparent, collaborative ecosystem.
1 Introduction
OpenSeeker addresses the scarcity of transparent, high-quality training data for search agents with two synthesis techniques and a fully open release. Using only SFT on 11.7k synthesized samples, it reports strong performance across several search benchmarks.
- Open-source search-agent development is constrained by proprietary training data, while existing open efforts often omit data, release only fragments, or lack competitive performance.
- OpenSeeker is presented as the first fully open-source search agent from a purely academic team to achieve frontier-level performance.The release includes the model, training data, QA pairs, and detailed trajectories.
- 11.7k synthesized samples and a single SFT run yield reported results of 29.5% on BrowseComp, 48.4% on BrowseComp-ZH, 74.0% on xbench-DeepSearch, and 59.4% item F1 on WideSearch.The reported setup uses only supervised fine-tuning, without heuristic filtering or hyperparameter optimization.
- 48.4% v.s. 46.7% on BrowseComp-ZH shows OpenSeeker surpassing Alibaba’s Tongyi DeepResearch on that benchmark.
- The proposed data pipeline combines fact-grounded scalable controllable QA synthesis with denoised trajectory synthesis.These techniques are designed to generate complex, multi-hop training tasks and higher-quality tool-use trajectories.
- The work fully open-sources its synthesis solution, final training dataset, and model weights to accelerate search-agent research.
2 Related Work
Search-agent research has largely followed proprietary development, leaving the community without the high-quality data needed to train competitive systems. OpenSeeker positions full pipeline and data release as a response to that gap.
- Contemporary search agents commonly use a ReAct reasoning-action-observation loop to interact with web environments.
- Corporate entities have historically dominated the development of high-performance search agents through proprietary systems and data.
- OpenSeeker addresses the data gap by releasing its entire synthesis pipeline and high-fidelity training data.
- The paper describes OpenSeeker as the first purely academic work to achieve state-of-the-art frontier-search benchmark performance while releasing full training data.
- The cited related-work footnote distinguishes OpenSeeker’s focus on data quality from concurrent efforts centered on context management.
3.1 Overview & Problem Formulation
OpenSeeker formulates deep-search training as constructing complex, fact-grounded questions and reliable long-horizon tool-use trajectories from the web graph. Its solution combines graph-based question construction with dynamic context denoising.
- The target dataset contains complex queries q, ground-truth answers y, and optimal tool-use trajectories τ*, intended to train long-horizon deep-search behavior.
- The web is modeled as a directed graph G = (V, E), with web pages as vertices and hyperlinks as edges.
- Sufficiently difficult questions must require multi-turn reasoning, tool calls, and tool responses rather than superficial retrieval.
- High-quality trajectories must be stable and reproducible so training signals reflect correct, generalizable strategies rather than stochastic successes.
- The proposed framework reverse-engineers an inference path in the web graph to construct questions that structurally require traversing it.Dynamic context denoising complements question synthesis by producing clearer reasoning and more precise tool calls.
3.2 Fact-Grounded Scalable Controllable QA Synthesis
The QA synthesis pipeline reverse-engineers web connectivity into fact-grounded, multi-hop questions, then filters them for strict tool necessity and logical solvability. Its design also supports scalable generation and deliberate control over reasoning complexity.
- Pipeline Overview: The pipeline constructs question-answer pairs directly from the web graph through generative construction and dual-criteria verification.It transforms hyperlinks into dynamic reasoning paths and filters candidates for difficulty and solvability.
- Generative Construction: Graph expansion samples a seed page and gathers connected outgoing nodes into a local dependency subgraph for question construction.The connected subgraph provides a topologically linked knowledge base.
- Generative Construction: Entity extraction condenses the expanded subgraph into key entities related to the seed page’s central theme.This reduces distracting content before question generation.
- Generative Construction: Questions are generated with a structural constraint requiring multiple entity-subgraph edges to derive the target answer, preventing simple lookup solutions.Entity and question obfuscation preserve the reasoning structure while replacing concrete entities with ambiguous descriptions.
- Dual-Criteria Verification: Candidates are retained only when a strong base model fails closed-book but succeeds with the entity subgraph as context.These criteria enforce external information seeking and logical consistency.
- Discussions: The framework provides factual grounding, scalable sample generation, and controllable difficulty through web topology and subgraph-size tuning.The authors report using approximately 68GB of English and 9GB of Chinese web data, while varying seed pages or graph configurations to produce diverse samples.
3.3 Denoised Trajectory Synthesis
Denoised trajectory synthesis separates clean teacher generation from raw-context student training. Retrospective summarization compresses older observations while preserving the latest raw response, and the student learns to produce teacher-derived reasoning and actions from noisy context.
- Motivation: The method decouples teacher generation from student training by using dynamic context denoising under web-scale context-window constraints.Raw observations are often dominated by irrelevant noise.
- Problem Formulation: A search trajectory contains a question, alternating reasoning, tool calls, and observations, followed by a final answer.The synthesis goal is to generate reasoning paths and actions that lead optimally to the answer.
- Retrospective Summarization: At each turn, the agent generates reasoning and an action from a context containing summarized long-term history and the immediately preceding raw observation.The protocol preserves complete recent information while maintaining concise long-term memory.
- Retrospective Summarization: After a new observation arrives, the previous observation is retrospectively compressed into a summary that replaces it in the long-term history.This rolling window filters noise and supports long horizons without performance degradation.
- Asymmetric Context Training: Teacher trajectories use denoised summaries, whereas student training uses the full raw context to supervise the teacher-derived reasoning and actions.The asymmetry is intended to make the final agent learn denoising and information extraction internally.
4 Experiments
OpenSeeker is evaluated against proprietary and open-source search agents across four benchmarks, including comparisons under SFT and comparable data volumes. It achieves strong performance with a small dataset, while its synthesized Chinese data is substantially more difficult than BrowseComp-ZH and its English data has comparable difficulty to BrowseComp-EN.
- Experimental Setup: OpenSeeker is evaluated on BrowseComp, BrowseComp-ZH, xbench-DeepSearch, and WideSearch, covering navigation, information location, deep research, and broad information retrieval.BrowseComp results use a 200-sample subset because of resource constraints.
- Experimental Setup: OpenSeeker is compared with proprietary models, large-scale open-source models, and other open-source SFT agents to assess performance across different resource regimes.The baseline groups include closed-source industry systems and large open-source models.
- Main Results: 11.7k training samples and standard SFT yield performance that rivals or exceeds resource-intensive corporate systems, including 48.4 on BrowseComp-ZH versus Tongyi DeepResearch’s 46.7.Tongyi DeepResearch uses CPT, SFT, and RL, whereas OpenSeeker uses only SFT.
- Main Results: 48.4 on BrowseComp-ZH exceeds WebSailor-V2-SFT’s 28.3 under the same SFT setup, while MiroThinker-32B-v0.1-SFT scores 13.8.The comparison is in the approximately 30B parameter class and attributes the advantage to data quality rather than data quantity alone.
- Main Results: 74.0 on xbench-DeepSearch and 59.4 item F1 on WideSearch exceed the best comparable-volume baseline combinations by nearly 8% and 15%, respectively.OpenSeeker uses approximately 11.7k samples versus baseline volumes of 10k–15k samples.
- Data Difficulty: OpenSeeker-v1-Data-EN exhibits difficulty comparable to BrowseComp-EN under inference with the same model.The English data has not yet been updated to the latest QA standards because of resource constraints.
- Data Difficulty: OpenSeeker-v1-Data-ZH averages 46.35 tool calls and 76.1k tokens per trajectory, versus 26.98 tool calls and 15.1k tokens for BrowseComp-ZH.The comparison uses the same inference model and indicates substantially higher difficulty for the synthesized Chinese data.
5 Discussions
OpenSeeker addresses the closed nature of high-performance search-agent development by fully open-sourcing a high-fidelity dataset and model. The authors present the current results as a lower bound and identify broader data, filtering, and tool integration as future directions.
- Breaking the Corporate Data Monopoly: OpenSeeker targets a search-agent ecosystem in which high-quality training data is a primary corporate advantage and existing open datasets often lack reasoning complexity.The discussion frames this as a long-standing data moat affecting academic development.
- Breaking the Corporate Data Monopoly: The fully open-sourced dataset is intended to provide resources for replicating and building upon industrial-grade search capabilities.The paper describes this contribution as breaking the long-standing data moat.
- Future Work: A single training run limits verification on more challenging data and exploration of parameters and data-filtering strategies.The authors describe the current work as a lower bound of OpenSeeker’s potential.
- Future Work: Future work will optimize data distributions, apply rigorous quality filtering, generate more complex training data, and integrate diverse tools and data sources.These directions aim to extend the agent beyond pure web search toward a more versatile framework.
6 Conclusions
OpenSeeker addresses the open-source search-agent data gap by releasing both high-fidelity synthesized training data and model weights. Using only 11.7k samples and one SFT run, it achieves state-of-the-art performance on frontier search benchmarks.
- OpenSeeker is the first purely academic search agent to open-source its complete training data and model weights while achieving state-of-the-art frontier-benchmark performance.The release includes synthesized data and weights, targeting the data monopoly that has constrained open-source search-agent research.
- 11.7k synthesized samples and a single supervised fine-tuning run produced the reported state-of-the-art results.The training used only 11.7k samples and one SFT run, contrasting with industrial baselines using more extensive resources and complex pipelines.
- Fact-grounded scalable controllable QA synthesis and denoised trajectory synthesis produced the high-fidelity training data used by OpenSeeker.The conclusion attributes data effectiveness to these two proposed synthesis methods.
- The complete dataset and model weights are released to support a more inclusive, transparent, and collaborative search-agent research ecosystem.
A Concurrent Works
OpenSeeker differs from concurrent open search-agent efforts in both data construction and openness of evaluation resources. Under the same SFT methodology, it outperforms OpenResearcher across three benchmarks and exceeds RedResearcher on BrowseComp-ZH.
- OpenResearcher aggregates QA pairs from existing open-source datasets and constructs trajectories in simulated environments, unlike OpenSeeker’s new graph-grounded QA and real-world web trajectories.
- Under the same SFT methodology, OpenSeeker comprehensively outperforms OpenResearcher across three benchmarks.
- 21.6% separates OpenSeeker from RedResearcher on BrowseComp-ZH, with scores of 48.4% versus 26.8%.
- OpenSeeker’s data requires significantly more tool calls than RedResearcher’s data when the same model is used for inference.