Deep search enables language agents to solve complex information-seeking tasks through iterative web exploration. However, developing robust open-source agents remains challenging for three reasons. First, opaque trajectory-construction pipelines make training data difficult to audit and reproduce. Second, context compression can create a substantial mismatch between training and inference. Third, agents may retrieve benchmark content from the open web during evaluation, contaminating the results and overstating their performance. To overcome these bottlenecks, we introduce Glasseek, a fully transparent deep search framework spanning the complete development lifecycle, releasing all data, model weights, training recipes, and code. Specifically, for data construction, we establish a verifiable trajectory curation pipeline that systematically generates high-quality exploration paths. For training, we propose Compaction-Managed GSPO (CM-GSPO), which uses the same policy for search and summarization while optimizing only execution sub-trajectories, enabling end-to-end search-policy training across context-compaction boundaries. For evaluation, we conduct the first systematic audit of test-set leakage in open-web agents and release a modular anti-leakage plugin that seamlessly integrates into existing evaluation harnesses. Under this rigorous evaluation protocol, our 9B agent outperforms frontier proprietary agents across almost all deep research benchmarks spanning diverse task types, and achieves the best overall performance among recent open-weight agents, setting a new state of the art at the 9B scale.
Figures & tables
Figure 1: Comparison of Glasseeker with representative deep-search agents and foundation models.
Figure 2: Overview of the High-Quality Data Construction Pipeline.
Figure 3: Overview of the CM-GSPO framework. Context compaction partitions each rollout into execution sub-trajectories with recorded conditioning contexts. CM-GSPO applies length-normalized sub-trajectory likelihood ratios and length-based advantage reweighting to correct the extra weighting introduced by independently averaging sub-trajectory losses. Summary-generation tokens are excluded from the RL objective.
Leaked Subset Acc. (%)
Overall Performance (%)
Model
Leakage Rate (%)
w/o Anti-Leakage
w/ Anti-Leakage
Base Acc.
Real Acc.
Drop
DeepSeek-V4-Flash
2.1
77.8
51.9
71.3
70.8
0.5
Apodex-mini
3.3
81.0
76.2
64.2
64.1
0.1
Agents-A1
11.2
91.6
46.5
63.4
58.3
5.1
XYZ-Aquila-mini
9.3
89.1
62.2
68.6
65.9
2.7
BigBang-V1
4.8
86.9
75.4
70.2
69.6
0.6
Table 1: External benchmark leakage analysis under different evaluation harnesses. We evaluate representative open-source deep-search agents that lead on benchmarks such as BrowseComp. Leakage Rate denotes the percentage of trajectories attempting to access benchmark-related artifacts. For the Leaked Subset (trajectories exhibiting leakage behavior), we report accuracy before ( w/o Anti-Leakage ) and after ( w/ Anti-Leakage ) harness-level defenses are applied. For the Overall dataset, we report accuracy before ( Base Acc. ) and after ( Real Acc. ) the same defenses, with Drop indicating the resulting overall accuracy decrease.
Model
BrowseComp
BrowseComp-ZH
GAIA-Text
XDS-2510
Seal-0
LiveBrowseComp
Closed-source LLMs
GPT-5.2
65.8
76.1
75.0
51.3
42.3
28.0
Claude-Opus-4.5
67.8
62.4
71.5
–
47.7
–
Gemini-3-Pro
59.2
60.2
–
71.6
51.5
51.0
Doubao-Seed-2.0-Pro
77.3
82.4
–
–
49.5
41.5
Open-Source Models ( ≥ 27B)
Table 2: Overall performance comparison (%) on search-oriented benchmarks. GAIA-Text denotes the text-only subset containing 103 questions, and XDS-2510 denotes the 2510 version of xbench-DeepSearch. Baseline scores marked with † are obtained using our evaluation harness; unmarked baseline scores are taken from the corresponding original reports and retain their original evaluation settings. All results for our models are obtained using our evaluation harness. Evaluation subsets for our runs are described in Section 5.1 . Bold numbers indicate the highest reported result in each column within each model group.
Method
BrowseComp
BrowseComp-ZH
Seal-0
XDS-2510
GAIA-Text
SFT Model (Baseline)
55.5
59.5
34.2
44.0
63.1
Vanilla GSPO
47.0 ( -8.5 )
51.2 ( -8.3 )
41.4 ( +7.2 )
47.0 ( +3.0 )
66.0 ( +2.9 )
+ Compaction
55.0 ( -0.5 )
55.7 ( -3.8 )
42.3 ( +8.1 )
48.0 ( +4.0 )
65.0 ( +1.9 )
+ CM-GSPO
61.6 ( +6.1 )
64.0 ( +4.5 )
50.0 ( +15.8 )
50.0 ( +6.0 )
75.7 ( +12.6 )
Table 3: Ablation of context compaction and CM-GSPO on the 9B model.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
SFT
RL
Training budget
3 epochs
100 steps
Learning rate
2×10−5
1×10−6
Minimum learning rate
2×10−6
–
Warmup ratio
0.03
–
Global batch size
64
16
Rollouts per prompt
–
8
Appendix
Table 4: Training hyperparameters and computational costs. Batch size denotes packed sequences for SFT and prompts for RL.
Figure 4: BrowseComp accuracy across attempt budgets. Markers show accuracy after one attempt and under the full retry strategy; colored segments show gains through three attempts and any further gains.
Figure 5: Topic distribution of SFT questions across three datasets. Topics are ordered clockwise by decreasing share in our data; bars within each topic represent Ours, QUEST, and OpenSeeker. Labels show within-dataset percentages. To keep the visualization compact, radial lengths above 20% use a threefold compression of the excess portion, indicated by double slashes. Bars start at 10% for Science & Technology and at zero otherwise; portions above 20% are compressed threefold, as indicated by double slashes.
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.
Training tool-use agents to improve from their own experience remains challenging, as supervised fine-tuning relies on fixed teacher-distilled trajectories, while sparse-reward reinforcement learning provides weak supervision for long-horizon interactions. We present DeepSearch-Evolve, a self-distillation framework for web agents built on DeepSearch-World, a deterministic and verifiable environment with reproducible search and page-reading tools. DeepSearch-World contains 420K multi-hop QA tasks constructed from entity-level random walks and supports key agentic cognitive behaviors useful for self-evolving, including progress verification, grounded reflection, and failure recovery. DeepSearch-Evolve iteratively performs trajectory generation, filtering, data mixing, and fine-tuning to train stronger agents. Without distillation from more capable models, DeepSearch-World-9B achieves competitive performance compared with open-source agents, reaching 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA, showing that verifiable environments enable scalable self-evolution for long-horizon web agents. We will release the environment, 420K training pool, validation set, model, and code to facilitate future research on self-improving deep search agents.
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.