Paper deep dive
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
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 98%
Last extracted: 3/22/2026, 5:24:30 AM
Summary
OpenSeeker is a fully open-source search agent developed by an academic team at Shanghai Jiao Tong University. It addresses the data scarcity in high-performance search agent development by introducing two technical innovations: fact-grounded scalable controllable QA synthesis (using web graph expansion and entity obfuscation) and denoised trajectory synthesis (using retrospective summarization). OpenSeeker achieves state-of-the-art performance among open-source agents on benchmarks like BrowseComp and xbench-DeepSearch, matching or exceeding proprietary industrial models.
Entities (5)
Relation Signals (3)
OpenSeeker â developedby â Shanghai Jiao Tong University
confidence 100% ¡ 1 Shanghai Jiao Tong University... we introduce OpenSeeker
OpenSeeker â outperforms â Tongyi DeepResearch
confidence 95% ¡ OpenSeeker... even surpasses industrial competitors such as Tongyi DeepResearch
OpenSeeker â trainedon â Qwen3-30B-A3B
confidence 95% ¡ perform Supervised Fine-Tuning (SFT) on the Qwen3-30B-A3B
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2603.15594v1
- Canonical: https://arxiv.org/abs/2603.15594v1
Trouble viewing inline? Open PDF directly â
Full Text
51,566 characters extracted from source content.
Expand or collapse full text
OpenSeeker: Democratizing Frontier Search Agents by Fully Open-Sourcing Training Data Yuwen Du 1,* , Rui Ye 1,*,#,â , Shuo Tang 1 , Xinyu Zhu 1 , Yijun Lu 1 , Yuzhu Cai 1 , Siheng Chen 1,â 1 Shanghai Jiao Tong University, * Equal Core Contributions, # Project Lead â Corresponding Authors: yr991129@sjtu.edu.cn, sihengc@sjtu.edu.cn Abstract 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. Codehttps://github.com/rui-ye/OpenSeeker Datahttps://huggingface.co/datasets/OpenSeeker/OpenSeeker-v1-Data Modelhttps://huggingface.co/OpenSeeker/OpenSeeker-v1-30B-SFT OpenSeeker-v1-30B-SFT DeepDive-32B MiroThinker-32B-DPO-v0.1 Tongyi-DeepResearch-30B WebSailor-V2-30B-RL WebSailor-V2-30B-SFT WebSailor-32B OpenAI-o3 GLM-4.6-357B DeepSeek-V3.1-671B Minimax-M2-230B Claude-4.5-Sonnet Claude-4-Sonnet 10 20 30 40 50 60 70 Score 48.4 29.7 17.0 46.7 44.1 28.3 25.5 68.7 49.5 49.2 48.5 42.4 22.5 ~30B Fully Open ~30B Mixed (Model-open/Closed) >30B or Closed-Source BrowseComp-ZH OpenSeeker-v1-30B-SFT DeepDive-32B MiroThinker-32B-DPO-v0.1 Tongyi-DeepResearch-30B WebSailor-V2-30B-RL WebLeaper-30B WebSailor-V2-30B-SFT WebSailor-32B OpenAI-o3 GLM-4.6-357B Minimax-M2-230B DeepSeek-V3.1-671B Claude-4.5-Sonnet Claude-4-Sonnet 10 20 30 40 50 29.5 15.3 13.0 43.4 35.3 27.7 24.4 10.5 49.1 45.1 44.0 30.0 24.1 14.7 ~30B Fully Open~30B Mixed (Model-open/Closed)>30B or Closed-Source BrowseComp OpenSeeker-v1-30B-SFT DeepDive-32B Tongyi-DeepResearch-30B WebSailor-V2-30B-RL WebLeaper-30B WebSailor-V2-30B-SFT WebSailor-32B DeepSeek-V3.1-671B OpenAI-o3 50 55 60 65 70 75 80 74.0 51.8 75.0 73.7 66.0 61.7 53.3 71.2 65.0 ~30B Fully Open ~30B Mixed (Model-open/Closed) >30B or Closed-Source xbench OpenSeeker-v1-30B-SFT WebLeaper-30B Claude-4-Sonnet OpenAI-o3 40 45 50 55 60 65 70 59.4 44.1 62.0 60.0 ~30B F/O ~30B Mixed >30B or Closed-Source WideSearch EN OpenSeeker (Ours, Fully Open-Source)Fully Open-SourceModel Open-SourceClosed-Source Figure 1: OpenSeeker stands out as the only fully open-source agent that achieves competitive performance on four search benchmarks, remarkably accomplishing this via simple SFT in a single training trial.. Preprint. arXiv:2603.15594v1 [cs.AI] 16 Mar 2026 1 Introduction In the era of information explosion, seeking accurate, real-time, and reliable information from the vast expanse of the internet has become a fundamental pillar of modern decision-making (Marchionini, 1995; Given et al., 2023). Consequently, the ability to perform deep search has emerged as a non-negotiable competency for frontier Large Language Model (LLM) agents (OpenAI, 2025a). The past year has witnessed a rapid rise in the development of search agents. As recently as April 10, 2025, even the most advanced LLMs, such as OpenAIâs o1 (OpenAI, 2024), struggled to surpass a score of 10 on the representative BrowseComp (Wei et al., 2025) benchmark. Yet, by March 2026, the landscape has shifted dramatically, with over ten agentic LLMs now exceeding the 50-point threshold (OpenAI, 2025b; Team et al., 2026a; Zeng et al., 2026), signaling a new era of autonomous web intelligence. However, despite this rapid progress, the training of high-performance search agents has remained a "closed- door game" played almost exclusively by well-funded corporate entities (OpenAI, 2026; Team et al., 2026a). The most capable search agents are currently dominated by proprietary models from giants such as Google and OpenAI. While prominent labs including Kimi and Minimax have contributed open-weights models, they have remained silent regarding their training data. Even within the research community, existing works either open-source the model without data (Li et al., 2025b), provide only a fraction of data (Li et al., 2025c), or fail to achieve competitive performance (Lu et al., 2025). This persistent lack of complete high-quality training data has stifled the growth of the open-source community for nearly a year. To bridge this gap, we, a purely academic team, introduce OpenSeeker, the first fully open-source search agent that achieves frontier-level performance in web search tasks. OpenSeeker is not merely an open- weights model; it is a comprehensive democratization of the search agent pipeline, providing the community with all of training data, including both complex question-answer (QA) pairs and detailed trajectories. The high-fidelity data behind OpenSeeker is powered by two core technical innovations: fact-grounded scalable controllable QA synthesis and denoised trajectory synthesis. Specifically, (1) our QA synthesis framework is designed to move beyond simple retrieval-based tasks that current models often solve through superficial pattern matching. To ensure queries demand genuine multi-hop reasoning, we reverse-engineer the web graph starting from randomly sampled seed pages within a massive web corpus. Specifically, we perform topological graph expansion to identify interconnected information clusters, which are then distilled into entity subgraphs. By applying entity obfuscation to these subgraphs, we transform straightforward facts into complex reasoning puzzles that structurally mandate multi-step navigation. This approach ensures our data isfact-grounded(anchored in real-world web topology),scalable(terabytes of web archives available), andcontrollable(modulating difficulty through subgraph complexity). (2) Our trajectory synthesis method is designed to overcome the distractions inherent in raw web content. During generation, we employ a secondary LLM to summarize preceding tool response, providing the teacher LLM with a cleaner/denoised history to produce superior reasoning and actions. In the training phase, however, we supervise the model to predict these expert decisions while conditioning it on the original, raw historical trajectory. This decoupling compels the agent to internalize robust information-extraction capabilities, learning to âsee through the noiseâ to identify the essential signals required for frontier-level performance. To validate the efficacy of our data, we synthesize a dataset comprising 10.3k English and 1.4k Chinese samples and perform Supervised Fine-Tuning (SFT) on the Qwen3-30B-A3B (Yang et al., 2025). Despite utilizing only SFT, OpenSeeker demonstrates remarkable competitiveness against models trained by corpo- rate entities across benchmarks including BrowseComp (Wei et al., 2025) (29.5%), BrowseComp-ZH (Zhou et al., 2025) (48.4%), xbench-DeepSearch (Xbench-Team, 2025) (74.0%), and WideSearch (Wong et al., 2025) (59.4% item F1) 1 . Notably, on the BrowseComp-ZH, OpenSeeker surpasses Alibabaâs Tongyi DeepResearch (Team et al., 2025d), a model trained with extensive continual pre-training, SFT and RL 1 It is worth highlighting that, due to resource constraints, these results are achieved in a single training run using default hyperparameters, without any heuristic filtering or hyperparameter optimization, leaving a large room for future research. 2 (48.4 v.s. 46.7). Among other models of equivalent scale trained via only SFT, our OpenSeeker achieves the bestperformance on average, proving the high-quality nature of our training data. Our primary contributions are summarized as follows: ⢠We propose two effective techniques: fact-grounded, scalable, controllable QA synthesis and denoised trajectory synthesis, enabling the automated generation of frontier-level training data. â˘We develop and release OpenSeeker, a search agent that achieves state-of-the-art performance among open-source agents, matching or exceeding frontier solutions developed by corporate. â˘We fully open-source the entire synthesis solution, the final training dataset (QA pairs and full trajectories), and the model weights, aiming to accelerating the development of search agents. Ultimately, to the best of our knowledge, OpenSeeker represents the first work by a purely academic team to achieve state-of-the-art performance on frontier search benchmarks while fully open-sourcing the entirety of its training data. Developed exclusively by an academic team, our work aims to democratize search intelligence by demonstrating that strategic data synthesis can effectively bridge the performance gap with industrial-scale efforts. By providing full data transparency, we hope OpenSeeker serves as a catalyst for the research community to participate in a more open, collaborative, and healthy development of autonomous agents. 2 Related Work The evolution of LLM-base search agents has shifted the paradigm of information retrieval from simple keyword matching to autonomous, multi-turn synthesis (Marchionini, 1995). Most contemporary search agents are architected upon the ReAct paradigm (Yao et al., 2023), which utilizes a reasoning-action- observation loop to interact with web environments 2 . Historically, this path has been dominated by corporate entities. (1) OpenAIâs Deep Research (OpenAI, 2025a) pioneers the fully closed-source path, followed by a series of proprietary agents including Kimi-Researcher (Kimi, 2025), Geminiâs Deep Research (DeepMind, 2025), and Perplexityâs Deep Research (Perplexity, 2025). (2) Within the past six months, a wave of "open-weights" models capable of search has emerged, such as the Kimi K2/2.5 series (Team et al., 2025b, 2026a), Zhipu GLM 4.5-5 (Zeng et al., 2025, 2026), MiniMax M2-2.5 (MiniMax, 2025, 2026), and Alibabaâs Tongyi DeepResearch (Team et al., 2025d). However, none of these industrial efforts have disclosed their training data, effectively maintaining a "data moat" that preserves frontier performance as a corporate secret. (3) While the research community has made significant strides with frameworks such as WebDancer (Wu et al., 2025), WebSailor (Li et al., 2025c), WebSailor-V2 (Li et al., 2025c), WebLeaper (Tao et al., 2025), AgentFounder (Su et al., 2026), DeepDive (Lu et al., 2025), and MiroThinker (MiroMind AI Team, 2025), they either lack public releases, provide only a small fraction of the data, or suffer from low data fidelity that fails to achieve competitive performance. This status quo has left the research community lacking of the high-quality data necessary to train high- performance agents. OpenSeeker explicitly addresses this void by fully open-sourcing its entire synthesis pipeline and high-fidelity training data, democratizing the "recipe" for frontier search intelligence 3 . To the best of our knowledge, OpenSeeker represents the first work by a purely academic team to achieve state-of-the-art performance on frontier search benchmarks while simultaneously open-sourcing the full training data. Notably, our SOTA results are achieved within a single training trial without any iterative refinement, underscoring the high quality of our synthesized data and leaving substantial room for future exploration. 2 While some parallel efforts focus on context management for agents (Ye et al., 2025; Team et al., 2025c), our work primarily focuses on the fundamental challenge of data quality. 3 We discuss with two concurrent works in Section A. 3 Web Corpus Graph Expansion Entity Extraction Question Generation Entity Obfuscation Choose Seed Question Obfuscation Difficulty CheckSolvability Check Final QA No No YesYes QA GenerationQA Verifier Figure 2: Overview of Fact-grounded scalable controllable QA synthesis. The pipeline begins with Graph Expansion, where a seed node is expanded into a subgraph of connected pages. Entity Extraction then distills key information themes into a structured Entity Subgraph. A generator synthesizes complex initial questions conditioned on this structure (Question Generation), ensuring multi-hop reasoning requirements. To enhance difficulty, we apply Entity Obfuscation to vagueify specific terms, finally producing a challenging question that necessitates deep graph traversal to solve. 3 Methodology 3.1 Overview & Problem Formulation Our primary objective is to synthesize a high-fidelity datasetD =(q,y,Ď â )comprising complex queries q, ground truth answersy, and optimal tool-use trajectoriesĎ â . This dataset aims to empower an agentĎ Î¸ to master long-horizon tool invocation for deep search tasks. We model the web as a directed graphG = (V,E ), whereVdenotes web pages andEdenotes hyperlinks. The synthesis challenge is to derive pairs(q,y)fromGsuch that solvingqnecessitates a trajectory Ď = [a 1 ,o 1 ,...,a T ,o T ]of lengthT ⍠1, wherea t are search actions ando t are observations. We argue that to effectively train deep search agents, one must address two pivotal challenges: (1) High-difficulty QA: Only sufficiently complex queries compel the system to engage in a rigorous multi-turn interaction cycle involving âReasoningâTool CallâTool Responseâ. This process is essential to generate long-horizon trajectories characterized by explicit decision points and extended tool invocation chains. (2) High-quality trajectories: The synthesis of solution paths must rely on stable and reproducible methods to ensure that the distilled training signals represent âcorrect and generalizableâ strategies rather than accidental successes derived from stochastic sampling. To address these, we propose a fact-grounded scalable controllable QA synthesis framework and a denoised trajectory synthesis method. The QA synthesis framework operates on the premise of reverse- engineering the reasoning graph: we first identify a latent inference path withinGand then construct a questionqthat structurally mandates traversing this path. Complementarily, our trajectory synthesis method utilizes dynamic context denoising to generate clear reasoning and precise tool calls. By subsequently training on raw trajectories, we enable the agent to intrinsically learn to denoise and extract relevant information from noisy tool responses. 3.2 Fact-Grounded Scalable Controllable QA Synthesis We engineer a pipeline to construct question-answer pairs(q,y)directly from the web graphG, as shown in Figure 2. By leveraging intrinsic connectivity, we transform static hyperlinks into dynamic reasoning paths, ensuring factual grounding and controllable complexity. This scalable framework operates in two distinct phases: Generative Construction to synthesize candidate pairs, and Dual-Criteria Verification to rigorously filter for difficulty and solvability. 4 3.2.1 Generative Construction: From Graph to Question Graph Expansion. To mimic the natural process of information discovery where one clue leads to another, we initiate the pipeline by sampling a seed nodev seed âź V. Recognising that complex questions rarely reside on a single isolated page, we expand fromv seed by traversing its outgoing edges inEto gather a set ofkconnected nodes. This forms a local dependency subgraphG sub =v seed âŞv i |(v seed ,v i )âE k , which serves as a coherent, topologically-linked knowledge base for problem construction. Entity Extraction. Synthesizing complex questions necessitates utilizing a generative model to reference the information cluster within the expanded subgraphG sub . However, the raw content of these nodes often contains excessive noise that can distract the generation model. To sharpen the focus, we identify the central themey theme ofv seed and execute an extraction function. This function distills a set of key entities from across the subgraph that are directly or indirectly related to the central themey theme , and reorganizes them into a condensed Entity SubgraphG entity . In this graph, nodes represent the extracted entities and edges preserve the original topological connections efficiently. This step effectively abstractsG sub into a dense relational structure, removing textual noise while retaining the essential logic paths. Question Generation. To prevent the generation of questions that can be solved by simple look-up, we employ a generatorP gen to synthesize an initial questionq init conditioned explicitly on the structure of the Entity SubgraphG entity . We impose a hard structural constraint: the derivation ofy theme fromq init must necessitate traversing multiple edges withinG entity . This explicitly forces the agent to engage in sequential multi-node deductive reasoning rather than single-step retrieval. Entity Obfuscation. The synthesized questions are intended to drive agents to perform multi-step ReAct reasoning. However, agents often exploit specific keywords to shortcut the reasoning process via direct search. To simulate realistic user ambiguity and dismantle these shortcuts, we apply an obfuscation operator ÎŚdirectly to the entity nodes inG entity . Concrete entitieseare mapped to vague, descriptive references Ěe = ÎŚ(e). This transformation yields a Fuzzy Entity Subgraph Ě G entity , where the structural connectivity remains intact but the semantic nodes now demand disambiguation. Question Obfuscation. The pipeline culminates in generating the final question Ěqby taking the initial questionq init and the fuzzy entity subgraph Ě G entity as inputs. This separation allows the generator to reference the pre-obfuscated descriptions in Ě G entity directly, thereby focusing exclusively on synthesizing the complex question structure. The generator rewritesq init to incorporate the ambiguous descriptions while preserving the original reasoning logic, with the target answer remaining the invariant y = y theme . 3.2.2 Dual-Criteria Verification via Rejection Sampling To ensure the synthesized pair( Ěq,y)is both challenging and valid, we employ a rejection sampling scheme based on two indicator functions: (1) Criterion 1: difficulty (strict tool necessity). LetĎ base be a strong foundation model. We define the difficulty condition asI[Ď base ( Ěq)̸= y], whereĎ base generates an answer in a closed-book setting (no external tools). If the model answers correctly using only parametric memory, the question is discarded. This guarantees that Ěq necessitates external information seeking. (2) Criterion 2:solvability (logical consistency).We define the solvability condition as I[Ď base ( Ěq|G entity ) = y]. Here, the model is provided with the full content of the Entity SubgraphG entity as context (oracle setting). If the model fails to derivey, it implies the reasoning path is broken or hallucinated. Such samples are rejected to strictly enforce logical validity. 3.2.3 Discussions Our data synthesis paradigm fundamentally advances agent training through three core strengths: 5 Reasoning Tool Call Reasoning Tool Call Reasoning Tool Call Raw Tool Response Reasoning Tool Call Raw Tool Response Summary ... Summarized Response Summarized Response Question Summarized Response Update Generate Reasoning Tool Call Raw Tool Response Reasoning Tool Call Raw Tool Response Reasoning Tool Call Raw Tool Response Reasoning Tool Call Raw Tool Response ... Question Training / Inference Denoised Trajectory Synthesis Generate expert-level reasoning and tool call conditioned on denoised history Learn to predict expert-level reasoning and tool call conditioned on raw history Figure 3: Overview of Denoised Trajectory Synthesis. We employ a retrospective summarization mech- anism where, after each tool call, the raw tool response from the previous turn is condensed into a âSummarized Responseâ that replaces the original raw tool response in the history window. This cleaner context enables the teacher to generate high-quality reasoning and actions. Note the asymmetry: while synthesis relies on summarized context, the training and inference phases operate on raw tool response to force the model to learn intrinsic denoising capabilities. (1) Factual grounding: By anchoring queries in the real webâs topology rather than relying on LLM generation, hallucination risks are significantly mitigated, if not entirely eliminated. Every training example is strictly grounded to verifiable, real-world data. (2) Scalability: In this work, we leverageâź68GB English andâź9GB Chinese web data to testify our solution, demonstrating that it suffices to synthesize high-quality QA pairs for training high-performance search agents. With TB-scale web archives still largely untapped, our pipeline transforms the open web into an inexhaustible source. By continuously varying seed pages or adjusting graph configurations, we can generate an (almost) infinite stream of diverse, non-repeating samples, ensuring no data bottlenecks for model scaling. (3) Controllability: In our solution, task difficulty is a deliberate design choice rather than a random variable By tuning the subgraph size (k), we can calibrate reasoning complexity and information coverage. This enables us to build tailored curriculums that progressively guide agents from straightforward retrieval to sophisticated, multi-hop investigations. 3.3 Denoised Trajectory Synthesis Constructing high-quality search trajectories requires strictly balancing information retention with context window constraints. In web-scale search, raw observations are often dominated by irrelevant noise. To address this, we propose a synthesis framework that technically decouples the generation context (Teacher) from the training context (Student), employing a dynamic context denoising strategy. 6 3.3.1 Problem Formulation Let a search trajectory be defined as a sequenceĎ = [q, (r 1 ,a 1 ,o 1 ),..., (r T ,a T ,o T ),y], whereqis the question,r t is the reasoning step (chain-of-thought),a t is the action (tool call), ando t is the observation (tool response) at turnt, culminating in the final answery. Our goal is to synthesize specific reasoning paths r t and actions a t that optimally lead to y. 3.3.2 Synthesis via Dynamic Context Denoising During trajectory synthesis, we employ a retrospective summarization mechanism. This ensures that the agent utilizes the complete information from the immediate past while maintaining a concise long-term memory. Formally, at turnt, the agent generates the reasoning and action pair(r t ,a t )based on the current contextH t . Our context construction follows a âSummarized History + Raw Recentâ protocol: H t =q, (r 1 ,a 1 ,s 1 ),..., (r tâ2 ,a tâ2 ,s tâ2 ) |z Summarized Long-Term History , (r tâ1 ,a tâ1 ,o tâ1 ) | z Raw Short-Term Context (1) wheres i = Summarize(o i |context)represents the compressed semantic summary of the observationo i . This mechanism operates in a two-phase cycle: (1) Decision phase (information usage): To generate the current decision(r t ,a t ), the agent is provided withH t , which includes the full raw observationo tâ1 from the immediately preceding step. This guarantees that the agent has access to all potential signals in the most recent observation to inform its next move, preventing premature information loss. (2) Compression phase (context denoising): Once steptis concluded and a new observationo t is obtained, the system retrospectively invokes a summarizer to compress the previous observationo tâ1 intos tâ1 . This summarys tâ1 then replaceso tâ1 in the long-term history for the next stepH t+1 . This rolling window approach effectively filters noise and denoises the context, enabling the generation of extremely long horizons without performance degradation. 3.3.3 Asymmetric Context Training for Robust Denoising To cultivate robustness in the final agent, we define a strategic asymmetry between the data format used for synthesis and that used for training, as shown in Figure 3. (1) Synthesis data (teacher): The trajectories are generated using the clean, denoised contextH t containing summaries. This acts as a scaffold, allowing the teacher model to produce âgoldenâ reasoning paths unencumbered by excessive noise. (2) Training data (student): For the final training dataset, we strip away the summaries and revert to the full raw context: H train t =q, (r 1 ,a 1 ,o 1 ),..., (r tâ1 ,a tâ1 ,o tâ1 )(2) The student model is supervised to predict the optimalr t ,a t (derived from the Teacher) given the noisy, raw contextH train t . This forces the student to implicitly learn the denoising and information extraction capabilities, effectively internalizing the context denoising logic within its own parameters to handle real-world unstructured data. 4 Experiments 4.1 Experimental Setup Implementation. We develop OpenSeeker, a deep search agent initialized from Qwen3-30B-A3B-Thinking- 2507 (Team, 2025), featuring 30B total parameters with 3B activated during prediction. The maximum tool call limit is set to 200, with any trajectory exceeding this threshold being forcibly terminated. The context window size is set to 256k. Each training sample comprises a user questionqand a sequence of 7 Table 1: Comparisons among our OpenSeeker and other search agents. â# Samplesâ denotes the number of total training data samples; â# OS Samplesâ denotes the number of open-source data samples; âTrainingâ denotes training techniques (CPT: continual pre-training, SFT: supervised fine-tuning, RL: reinforcement learning); âAcademicâ denotes whether conducted by pure academic team (â: Yes,Ă: No); âBC-ZHâ denotes BrowseComp-ZH; âWideSearchâ denotes item F1 result on the English subset. With simple SFT, OpenSeeker-v1-30B-SFT even surpasses Tongyi DeepResearch on BrowseComp-ZH which is trained via CPT, SFT, and RL. Among all SFT-based agents, OpenSeeker performs significantly best. Model Name# Samples# OS SamplesTrainingAcademicBrowseCompBC-ZHxbenchWideSearch Closed-Source Proprietary Models Claude-4-Sonnet?0?Ă14.722.5-62.0 Claude-4.5-Sonnet?0?Ă24.142.4-- Claude-4-Opus?0?Ă18.837.4-- OpenAI-o3?0?Ă49.168.7-60.0 OpenAI Deep Research?0?Ă51.542.9-- GPT-5-High?0?Ă54.963.0-- Open-Source Models > 30B Kimi-K2-Instruct-1T?0?Ă14.128.8-59.9 DeepSeek-V3.1-671B?0?Ă30.049.271.2- DeepSeek-V3.2-671B?0?Ă51.465.0-- GLM-4.6-357B?0?Ă45.149.5-- GLM-4.7-357B?0?Ă52.066.6-- Minimax-M2-230B?0?Ă44.048.5-- âź30B Models WebDancer-32B?0SFT + RLĂ3.818.0-- MiroThinker-32B-v0.1147 k147 kSFTĂ10.613.8-- MiroThinker-32B-v0.1147 k147 kSFT + RLĂ13.017.0-- DeepDive-32B4.1 k4.1 kSFT + RLĂ15.329.751.8- WebSailor-32B?0SFT + RLĂ10.525.553.3- WebSailor-V2-30B?0SFTĂ24.428.361.7- WebSailor-V2-30B?0SFT + RLĂ35.344.173.7- WebLeaper-30B15 k0SFTĂ27.7-66.044.1 Tongyi DeepResearch?0CPT + SFT + RLĂ43.446.775.0- OpenSeeker-v1-30B-SFT11.7 k11.7 kSFTâ29.548.474.059.4 raw reasoning steps, tool calls, and full, uncompressed tool responses. Note that due to resource constraints, we only train the model for a single run, without any heuristic data filtering or hyperparameter-tuning for training, leaving a large improving room for future research. Benchmarks. We evaluate OpenSeeker on four key benchmarks: BrowseComp (Wei et al., 2025) and BrowseComp-ZH (Zhou et al., 2025), which test multi-step navigation and hard information location in English and Chinese, respectively (we evaluate BrowseComp results on a subset of 200 samples due to resource constraints); xbench-DeepSearch (Xbench-Team, 2025), assessing complex deep research capabilities like planning and synthesis; and WideSearch (Wong et al., 2025), measuring reliability in broad information seeking across extensive sources. Baselines. To assess the efficacy of OpenSeeker, we compare it against a broad spectrum of state-of-the-art systems categorized into three groups: (1) closed-source proprietary models, representing the industry upper bound (e.g., Claude series (Anthropic, 2025), OpenAI-o3 (OpenAI, 2025b), OpenAI Deep Re- search (OpenAI, 2025a), GPT-5-High (Singh et al., 2025)); (2) large-scale open-source models, comprising massive parameter systems such as Kimi-K2 (Team et al., 2025b), DeepSeek (V3.1/V3.2 (DeepSeek- AI et al., 2025)), GLM-4 (4.6/4.7) (Team et al., 2025a), Minimax-M2 (MiniMax AI Team, 2025), and LongCat-Flash (Team et al., 2026b); and (3)âź30B models, which serve as direct, comparable-scale benchmarks. This group includes representative search agents such as MiroThinker series (MiroMind AI Team, 2025), DeepDive-32B (Lu et al., 2025), WebDancer (Wu et al., 2025), WebSailor-V2 (Li et al., 2025b), WebLeaper (Tao et al., 2025), and Tongyi DeepResearch (Li et al., 2025a). Baseline performance metrics are derived from their respective technical reports or public leaderboards. 8 Table 2: Performance comparison of different models trained via SFT. Our OpenSeeker consistently and significantly performs the best across four benchmarks with only 11.7k training samples. Data# Samples# OS SamplesAcademicBrowseCompBC-ZHxbenchWideSearch-EN DeepDive-32B4.1 k4.1 kĂ9.523.048.5- MiroThinker-32B-v0.1147 k147 kĂ10.613.8-- WebSailor-V2-30B?0Ă24.428.361.7- WebLeaper-30B15 k0Ă27.7-66.044.1 OpenSeeker-v1-30B-SFT11.7 k11.7kâ29.548.474.059.4 Table 3: Performance comparison under comparable data volumes. OpenSeeker achieves significant advantages across three benchmarks, demonstrating the high quality of our data. Data# Samples# OS SamplesDeveloperBrowseCompxbenchWideSearch-EN WebSailor-V2-10k10k0Tongyi24.5062.6738.91 WebSailor-V2-5k + WebLeaper-Basic-5k10k0Tongyi20.6758.3332.26 WebSailor-V2-5k + WebLeaper-Union-5k10k0Tongyi27.5062.3341.70 WebSailor-V2-5k + WebLeaper-Reverse-Union-10k15k0Tongyi27.6766.0044.07 OpenSeeker-v1-Data-11.7k11.7 k11.7kAcademic29.5074.0059.40 4.2 Main Results Outperforming resource-intensive industry baselines. As shown in Table 1, our primary evaluation compares OpenSeeker against a spectrum of proprietary and open-source models, highlighting its superior efficiency and competitiveness achieved through just a single training run. Despite utilizing a modest dataset of only 11.7k samples and a standard SFT protocol, OpenSeeker consistently rivals or exceeds the performance of models backed by massive corporate resources. A standout result is observed on the BrowseComp-ZH benchmark, where OpenSeeker achieves a score of 48.4, surpassing Alibabaâs Tongyi DeepResearch (46.7). This is particularly significant given that Tongyi DeepResearch employs a complex, resource-heavy training pipeline involving Continual Pre-Training (CPT), SFT, and Reinforcement Learning (RL), whereas OpenSeeker relies solely on high-quality SFT data. Superior performance under identical training setup. As shown in Table 2, when evaluated under the same SFT training protocol in theâź30B parameter class, OpenSeeker demonstrates a decisive advantage, highlighting the effectiveness of our data synthesis method. Most notably on BrowseComp-ZH, OpenSeeker (48.4) outperforms the runner-up WebSailor-V2-SFT (28.3) by nearly 20%. Additionally, models such as MiroThinker-32B-v0.1-SFT (13.8) lag significantly behind, confirming that data quantity (e.g., 147k for MiroThinker) is secondary to data quality. Our Denoised Trajectory Synthesis effectively teaches the model to denoise and extract âneedle-in-a-haystackâ information from raw, noisy web observation, a capability that standard SFT datasets often fail to cultivate. Superior performance with comparable data volume. To further isolate the contribution of our synthesis methodology, we compare OpenSeeker against various configurations of WebSailor-V2 and WebLeaper data. As shown in Table 3, despite using a comparable or even smaller volume of data (âź11.7k samples vs. 10kâ15k samples), OpenSeeker demonstrates superior performance across all benchmarks. Specifically, on xbench and WideSearch, it outperforms the best baseline combination (utilizing 15k samples) by nearly 8% (74.0) and 15% (59.4), respectively. This result strongly validates the high quality and efficiency of our data, demonstrating that our synthesized samples provide significantly more effective supervision signals. In stark contrast to baselines relying on proprietary, company-synthesized datasets, OpenSeeker, developed by a purely academic research team, achieves this efficiency by leveraging our independently synthesized high-difficulty QA and high-quality denoised trajectories, while fully open-sourcing the entire dataset to the community. 9 Figure 4: Comparison of difficulty between OpenSeeker-v1-Data-ZH and BrowseComp-ZH using the same model for inference. OpenSeeker-v1-Data-ZH exhibits significantly higher average token counts and tool call counts than BrowseComp-ZH. Figure 5: Comparison of difficulty between OpenSeeker-v1-Data-EN and BrowseComp-EN using the same model for inference. OpenSeeker-v1-Data-EN exhibits difficulty comparable to that of BrowseComp-EN. Data statistics analysis. To quantitatively contrast the difficulty of our synthesized data with that of standard benchmarks, we employ an open-source model to perform inference on both our synthesized samples and the BrowseComp benchmarks (including BrowseComp-ZH and BrowseComp-EN). The comparison reveals that our synthesized data matches or even exceeds the difficulty of established benchmarks. Notably, although our Chinese dataset contains only approximately 1.4k samples, its complexity significantly surpasses that of BrowseComp-ZH. As illustrated in Figure 4, our synthesized Chinese data averages 46.35 tool calls per trajectory with an average token length of 76.1k, whereas BrowseComp-ZH averages only 26.98 tool calls and 15.1k tokens. This not only demonstrates that our problems are inherently more challenging but also validates that despite the limited data volume, its high fidelity and complexity directly contribute to superior performance on Chinese benchmarks. Due to resource constraints, our English data has not yet been updated to the latest QA standards, resulting in slightly lower difficulty compared to the Chinese data (see Figure 5). We expect to release an updated version in the near future. 5 Discussions Breaking the corporate data monopoly. For a long period, the development of high-performance search agents has been a âclosed-door gameâ dominated by tech corporate, with high-quality data serving as their primary moat. Concurrently, existing open-source datasets often suffer from poor quality and inadequate reasoning complexity, leaving the academic community ill-equipped to train truly capable, frontier-level search models. OpenSeeker addresses this critical bottleneck. By open-sourcing a high-fidelity dataset that 10 enables frontier-level performance, we provide the community with the necessary resources to replicate and build upon industrial-grade capabilities, breaking the long-standing âdata moatâ. Future work. While our current work demonstrates significant effectiveness, it represents merely a lower bound of OpenSeekerâs potential. Due to resource constraints, we can only train for a single run, limiting not only the verification of effectiveness on more challenging data, but also the exploration of various parameters and data filtering strategies. In the next phase, we aim to optimize data distributions, implement rigorous quality filtering, and generate training data of even higher complexity to push the boundaries of performance. Furthermore, we plan to extend the agentâs capabilities beyond pure web search by integrating a more diverse set of tools and data sources, ultimately advancing toward a more versatile and generalist agentic framework. 6 Conclusions The growth of the open-source search agent community has long been stifled by the monopoly of high- quality training data held by industrial corporations. To bridge this gap, OpenSeeker represents the first work by a purely academic team to achieve state-of-the-art performance on frontier search benchmarks while simultaneously open-sourcing the full training data. Notably, our SOTA results are achieved using only 11.7k synthesized samples through a single supervised fine-tuning run, surpassing industrial baselines that rely on extensive resources and complex training pipelines. This efficiency validates the effectiveness of our proposed fact-grounded scalable controllable QA synthesis and denoised trajectory synthesis methods in producing high-fidelity training data. By openly releasing our the complete dataset, and model weights, we aim to dismantle the data barriers in this domain and foster a more inclusive, transparent, and collaborative ecosystem for future search agent research. References Anthropic. Introducing claude 4, 2025. URL https://w.anthropic.com/news/claude-4. Zheng Chu, Xiao Wang, Jack Hong, Huiming Fan, Yuqi Huang, Yue Yang, Guohai Xu, Chenxiao Zhao, Cheng Xiang, Shengchao Hu, et al. Redsearcher: A scalable and cost-efficient framework for long- horizon search agents. arXiv preprint arXiv:2602.14234, 2026. Google DeepMind.Gemini 2.5, 2025.URLhttps://blog.google/technology/ google-deepmind/gemini-model-thinking-updates-march-2025/. DeepSeek-AI, Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenhao Xu, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Erhang Li, Fangqi Zhou, Fangyun Lin, Fucong Dai, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Hanwei Xu, Hao Li, Haofen Liang, Haoran Wei, Haowei Zhang, Haowen Luo, Haozhe Ji, Honghui Ding, Hongxuan Tang, Huanqi Cao, Huazuo Gao, Hui Qu, Hui Zeng, et al. Deepseek-v3.2: Pushing the frontier of open large language models. arXiv preprint arXiv:2512.02556, December 2025. doi: 10.48550/arXiv.2512.02556. URL https://arxiv.org/abs/2512.02556. Lisa M Given, Donald O Case, and Rebekah Willson. Looking for information: Examining research on how people engage with information. Emerald Publishing Limited, 2023. Kimi. Kimi-researcher: End-to-end rl training for emerging agentic, 2025. URLhttps://moonshotai. github.io/Kimi-Researcher/. 11 Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, Kuan Li, Liangcai Su, Litu Ou, Liwen Zhang, Pengjun Xie, Rui Ye, Wenbiao Yin, Xinmiao Yu, Xinyu Wang, Xixi Wu, Xuanzhong Chen, Yida Zhao, Zhen Zhang, Zhengwei Tao, Zhongwang Zhang, Zile Qiao, Chenxi Wang, Donglei Yu, Gang Fu, Haiyang Shen, Jiayin Yang, Jun Lin, Junkai Zhang, Kui Zeng, Li Yang, Hailong Yin, Maojia Song, Ming Yan, Minpeng Liao, Peng Xia, Qian Xiao, Rui Min, Ruixue Ding, Runnan Fang, Shaowei Chen, Shen Huang, Shihang Wang, Shihao Cai, Weizhou Shen, Xiaobin Wang, Xin Guan, Xinyu Geng, Yingcheng Shi, Yuning Wu, Zhuo Chen, Zijian Li, and Yong Jiang. Tongyi deepresearch technical report. arXiv preprint arXiv:2510.24701, Oct 2025a. doi: 10.48550/arXiv.2510.24701. Kuan Li, Zhongwang Zhang, Huifeng Yin, Rui Ye, Yida Zhao, Liwen Zhang, Litu Ou, Dingchu Zhang, Xixi Wu, Jialong Wu, et al. Websailor-v2: Bridging the chasm to proprietary agents via synthetic data and scalable reinforcement learning. arXiv preprint arXiv:2509.13305, 2025b. Kuan Li, Zhongwang Zhang, Huifeng Yin, Liwen Zhang, Litu Ou, Jialong Wu, Wenbiao Yin, Baixuan Li, Zhengwei Tao, Xinyu Wang, et al. Websailor: Navigating super-human reasoning for web agent. arXiv preprint arXiv:2507.02592, 2025c. Zhuofeng Li, Dongfu Jiang, Xueguang Ma, Haoxiang Zhang, Ping Nie, Yuyu Zhang, Kai Zou, Jianwen Xie, Yu Zhang, and Wenhu Chen.Openresearcher:A fully open pipeline forlong-horizondeepresearchtrajectorysynthesis.https://w.notion.so/ OpenResearcher-A-Fully-Open-Pipeline-for-Long-Horizon-Deep-Research-Trajectory-Synthesis-2f7e290627b5800cb3a0cd7e8d6ec0ea, 2025d. Notion Blog. Rui Lu, Zhenyu Hou, Zihan Wang, Hanchen Zhang, Xiao Liu, Yujiang Li, Shi Feng, Jie Tang, and Yuxiao Dong. Deepdive: Advancing deep search agents with knowledge graphs and multi-turn rl. arXiv preprint arXiv:2509.10446, 2025. Gary Marchionini. Information seeking in electronic environments. Number 9. Cambridge university press, 1995. MiniMax. Minimax m2 & agent: Ingenious in simplicity, 2025. URLhttps://w.minimax.io/ news/minimax-m2. MiniMax. Minimax m2.5: Built for real-world productivity, 2026. URLhttps://w.minimax.io/ news/minimax-m25. MiniMax AI Team. MiniMax M2 & Agent: Ingenious in Simplicity, October 2025. URLhttps:// w.minimax.io/news/minimax-m2. Open-sourced model weights on Hugging Face:https: //huggingface.co/MiniMaxAI/MiniMax-M2. MiroMind AI Team. Mirothinker: An open-source agentic model series trained for deep research and complex, long-horizon problem solving, 2025. URLhttps://github.com/MiroMindAI/ MiroThinker. OpenAI.Introducingopenaio1-preview.https://openai.com/index/ introducing-openai-o1-preview/, 2024. Accessed: 2025-01-22. OpenAI.Deep research system card,2025a.URLhttps://cdn.openai.com/ deep-research-system-card.pdf. OpenAI.Introducing openai o3 and o4-mini, 2025b.URLhttps://openai.com/index/ introducing-o3-and-o4-mini/. OpenAI.Introducinggpt-5.2,2026.URLhttps://openai.com/index/ introducing-gpt-5-2/. 12 Perplexity. Introducing perplexity deep research, 2025. URLhttps://w.perplexity.ai/hub/ blog/introducing-perplexity-deep-research. Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, Akshay Nathan, Alan Luo, Alec Helyar, Aleksander Madry, Aleksandr Efremov, Aleksandra Spyra, Alex Baker-Whitcomb, Alex Beutel, Alex Karpenko, Alex Makelov, Alex Neitz, et al. Openai gpt-5 system card. arXiv preprint arXiv:2601.03267, December 2025. doi: 10.48550/arXiv.2601.03267. URL https://arxiv.org/abs/2601.03267. Liangcai Su, Zhen Zhang, Guangyu Li, Zhuo Chen, Chenxi Wang, Maojia Song, Xinyu Wang, Kuan Li, Jialong Wu, Xuanzhong Chen, Zile Qiao, Zhongwang Zhang, Huifeng Yin, Shihao Cai, Runnan Fang, Zhengwei Tao, Wenbiao Yin, Rui Ye, Yong Jiang, Ningyu Zhang, Pengjun Xie, Fei Huang, Kai Ye, Kewei Tu, Chenxiong Qian, and Jingren Zhou. Scaling agents via continual pre-training. In The Fourteenth International Conference on Learning Representations, 2026. URLhttps://openreview.net/ forum?id=Dru5m9anE. Zhengwei Tao, Haiyang Shen, Baixuan Li, Wenbiao Yin, Jialong Wu, Kuan Li, Zhongwang Zhang, Huifeng Yin, Rui Ye, Liwen Zhang, et al. Webleaper: Empowering efficiency and efficacy in webagent via enabling info-rich seeking. arXiv preprint arXiv:2510.24697, 2025. GLM Team, Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, Kedong Wang, Lucen Zhong, Mingdao Liu, Rui Lu, Shulin Cao, Xiaohan Zhang, Xuancheng Huang, Yao Wei, Yean Cheng, Yifan An, Yilin Niu, Yuanhao Wen, Yushi Bai, Zhengxiao Du, Zihan Wang, Zilin Zhu, Bohan Zhang, Bosi Wen, Bowen Wu, Bowen Xu, Can Huang, Casey Zhao, Changpeng Cai, Chao Yu, Chen Li, Chendi Ge, Chenghua Huang, Chenhui Zhang, Chenxi Xu, Chenzheng Zhu, Chuang Li, Congfeng Yin, Daoyan Lin, Dayong Yang, Dazhi Jiang, Ding Ai, Erle Zhu, Fei Wang, Gengzheng Pan, Guo Wang, Hailong Sun, Haitao Li, Haiyang Li, Haiyi Hu, Hanyu Zhang, Hao Peng, Hao Tai, Haoke Zhang, Haoran Wang, Haoyu Yang, He Liu, He Zhao, Hongwei Liu, Hongxi Yan, Huan Liu, Huilong Chen, Ji Li, Jiajing Zhao, Jiamin Ren, Jian Jiao, Jiani Zhao, Jianyang Yan, Jiaqi Wang, Jiayi Gui, Jiayue Zhao, Jie Liu, Jijie Li, Jing Li, Jing Lu, Jingsen Wang, Jingwei Yuan, Jingxuan Li, Jingzhao Du, Jinhua Du, Jinxin Liu, Junkai Zhi, Junli Gao, Ke Wang, Lekang Yang, Liang Xu, Lin Fan, Lindong Wu, Lintao Ding, Lu Wang, Man Zhang, Minghao Li, Minghuan Xu, Mingming Zhao, Mingshu Zhai, Pengfan Du, Qian Dong, Shangde Lei, Shangqing Tu, Shangtong Yang, Shaoyou Lu, Shijie Li, Shuang Li, Shuang-Li, Shuxun Yang, Sibo Yi, Tianshu Yu, Wei Tian, Weihan Wang, Wenbo Yu, Weng Lam Tam, Wenjie Liang, Wentao Liu, Xiao Wang, Xiaohan Jia, Xiaotao Gu, Xiaoying Ling, Xin Wang, Xing Fan, Xingru Pan, Xinyuan Zhang, Xinze Zhang, Xiuqing Fu, Xunkai Zhang, Yabo Xu, Yandong Wu, Yida Lu, Yidong Wang, Yilin Zhou, Yiming Pan, Ying Zhang, Yingli Wang, Yingru Li, Yinpei Su, Yipeng Geng, Yitong Zhu, Yongkun Yang, Yuhang Li, Yuhao Wu, Yujiang Li, Yunan Liu, Yunqing Wang, Yuntao Li, Yuxuan Zhang, Zezhen Liu, Zhen Yang, Zhengda Zhou, Zhongpei Qiao, Zhuoer Feng, Zhuorui Liu, Zichen Zhang, Zihan Wang, Zijun Yao, Zikang Wang, Ziqiang Liu, Ziwei Chai, Zixuan Li, Zuodong Zhao, Wenguang Chen, Jidong Zhai, Bin Xu, Minlie Huang, Hongning Wang, Juanzi Li, Yuxiao Dong, and Jie Tang. Glm-4.5: Agentic, reasoning, and coding (arc) foundation models, 2025a. URL https://arxiv.org/abs/2508.06471. Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, et al. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534, 2025b. Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, SH Cai, Yuan Cao, Y Charles, HS Che, Cheng Chen, Guanduo Chen, et al. Kimi k2. 5: Visual agentic intelligence. arXiv preprint arXiv:2602.02276, 2026a. Meituan LongCat Team, Anchun Gui, Bei Li, Bingyang Tao, Bole Zhou, Borun Chen, Chao Zhang, Chen Gao, Chen Zhang, Chengcheng Han, et al. Longcat-flash-thinking-2601 technical report, 2026b. 13 MiroMind Team, Song Bai, Lidong Bing, Carson Chen, Guanzheng Chen, Yuntao Chen, Zhe Chen, Ziyi Chen, Jifeng Dai, Xuan Dong, et al. Mirothinker: Pushing the performance boundaries of open-source research agents via model, context, and interactive scaling. arXiv preprint arXiv:2511.11793, 2025c. Qwen Team. Qwen3-30b-a3b-thinking-2507, 2025. URLhttps://huggingface.co/Qwen/ Qwen3-30B-A3B-Thinking-2507. Tongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, et al. Tongyi deepresearch technical report. arXiv preprint arXiv:2510.24701, 2025d. Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. Browsecomp: A simple yet challenging benchmark for browsing agents. arXiv preprint arXiv:2504.12516, 2025. Ryan Wong, Jiawei Wang, Junjie Zhao, Li Chen, Yan Gao, Long Zhang, Xuan Zhou, Zuo Wang, Kai Xiang, Ge Zhang, et al. Widesearch: Benchmarking agentic broad info-seeking. arXiv preprint arXiv:2508.07999, 2025. Jialong Wu, Baixuan Li, Runnan Fang, Wenbiao Yin, Liwen Zhang, Zhengwei Tao, Dingchu Zhang, Zekun Xi, Yong Jiang, Pengjun Xie, et al. Webdancer: Towards autonomous information seeking agency. arXiv preprint arXiv:2505.22648, 2025. Xbench-Team. Xbench-deepsearch, 2025. URL https://xbench.org/agi/aisearch. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. Re- act: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023. Rui Ye, Zhongwang Zhang, Kuan Li, Huifeng Yin, Zhengwei Tao, Yida Zhao, Liangcai Su, Liwen Zhang, Zile Qiao, Xinyu Wang, et al. Agentfold: Long-horizon web agents with proactive context management. arXiv preprint arXiv:2510.24699, 2025. Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, et al. Glm-4.5: Agentic, reasoning, and coding (arc) foundation models. arXiv preprint arXiv:2508.06471, 2025. Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chengx- ing Xie, Cunxiang Wang, et al. Glm-5: from vibe coding to agentic engineering. arXiv preprint arXiv:2602.15763, 2026. Peilin Zhou, Bruce Leon, Xiang Ying, Can Zhang, Yifan Shao, Qichen Ye, Dading Chong, Zhiling Jin, Chenxuan Xie, Meng Cao, et al. Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese. arXiv preprint arXiv:2504.19314, 2025. 14 Table 4: Performance comparison of concurrent works trained via SFT. Data# Samples# OS SamplesAcademicBrowseCompBC-ZHxbench Openresearcher96k96kĂ26.3-65 REDSearcher?10kĂ34.726.8- OpenSeeker-v1-Data-11.7k11.7 k11.7kâ29.548.474.0 Figure 6: Comparison of tool call counts using the same model for inference: OpenSeekerâs data vs. REDSearcherâs data. OpenSeekerâs data demonstrates a significantly higher average number of tool calls. A Concurrent Works As the field moves toward more transparent search agent development, several concurrent efforts have emerged, yet they differ significantly from OpenSeeker in methodology and openness. (1) OpenRe- searcher (Li et al., 2025d) primarily aggregates QA pairs from existing open-source datasets and constructs trajectories within simulated environments. In contrast, OpenSeeker creates entirely new, high-difficulty QA pairs via our graph-grounded synthesis and collects trajectories within real-world web environments to ensure better generalizability. Furthermore, OpenSeeker demonstrates superior data quality, outper- forming OpenResearcher (that is trained using 96k samples) with only 11.7k high-fidelity samples. (2) RedResearcher (Chu et al., 2026) adopts a multi-stage pipeline involving mid-training, SFT, and Reinforce- ment Learning (RL). However, it lacks full transparency regarding its training protocol and only provides a partial release of its SFT and RL data. Crucially, both OpenResearcher and RedResearcher involve significant corporate participation. OpenSeeker distinguishes itself as the first academic-led initiative to achieve state-of-the-art performance with a lean, SFT-only approach and 100% data transparency, proving that strategic data synthesis can bridge the gap traditionally filled by massive corporate compute and iterative RL cycles. Quantitative results further validate these advantages. As shown in Table 4, when trained using the same SFT methodology, OpenSeeker comprehensively outperforms OpenResearcher across three benchmarks. Notably, on BrowseComp-ZH, OpenSeeker surpasses RedResearcher by 21.6% (48.4% vs 26.8%). Fur- thermore, as illustrated in Figure 6, when employing the same model for inference to compare tool call counts, OpenSeekerâs data proves significantly more challenging than that of RedResearcher. Specifi- cally, the average number of tool calls for OpenSeeker-v1-Data-EN averages 45.92 calls against 36.91 for RedResearcher-EN, while OpenSeeker-v1-Data-ZH is 46.35 compared to 20.02 for RedResearcher-ZH. 15