Source-linked AI summary
WebSailor-V2: Bridging the Chasm to Proprietary Agents via Synthetic Data and Scalable Reinforcement Learning
Kuan Li, Zhongwang Zhang, Huifeng Yin, Rui Ye, Yida Zhao, Liwen Zhang, Litu Ou, Dingchu Zhang, Xixi Wu, Jialong Wu, Xinyu Wang, Zile Qiao, Zhen Zhang, Yong Jiang, Pengjun Xie, Fei Huang, Jingren Zhou
TL;DR
WebSailor-V2 targets the gap between open-source and proprietary deep research agents on complex information-seeking tasks. It combines high-uncertainty synthetic data, SFT cold start, controlled environments, dynamic data curation, and agentic RL, achieving strong cross-benchmark performance that rivals proprietary systems and surpasses prior open-source agents.
Problem
Open-source agents lag proprietary systems on complex information-seeking tasks, partly because their data lacks broad logical structures and diverse uncertainty types.
Method
WebSailor-V2 combines SailorFog-QA-V2 data construction with SFT, simulated and real RL environments, automated data curation, and a tailored GRPO-based training strategy.
Results
WebSailor-V2 achieves state-of-the-art results among open-source agents across web-agent benchmarks and is highly competitive with top-tier proprietary agents.
Takeaways & Limitations
A 30B-sized WebSailor-V2 agent can rival advanced proprietary Deep Research agents while significantly surpassing previous open-source solutions.
Abstract
from arXiv · showhide
Transcending human cognitive limitations represents a critical frontier in LLM training. Proprietary agentic systems like DeepResearch have demonstrated superhuman capabilities on extremely complex information-seeking benchmarks such as BrowseComp, a feat previously unattainable. We posit that their success hinges on a sophisticated reasoning pattern absent in open-source models: the ability to systematically reduce extreme uncertainty when navigating vast information landscapes. Based on this insight, we introduce WebSailor, a complete post-training methodology designed to instill this crucial capability. Our approach involves generating novel, high-uncertainty tasks through structured sampling and information obfuscation, RFT cold start, and an efficient agentic RL training algorithm, Duplicating Sampling Policy Optimization (DUPO). With this integrated pipeline, WebSailor significantly outperforms all open-source agents in complex information-seeking tasks, matching proprietary agents' performance and closing the capability gap.
1 Introduction
WebSailor-V2 addresses the open-source performance gap in deep research agents through a complete post-training pipeline spanning data construction, SFT, and reinforcement learning. Its enhanced data and training strategy produces strong benchmark performance, including results surpassing a much larger open-source agent.
- Open-source deep research agents still face a substantial performance gap on complex information-seeking tasks.
- Insufficient data diversity and narrow uncertainty definitions limit agents’ ability to generalize across complex logical structures.
- WebSailor-V2 introduces SailorFog-QA-V2, with improved knowledge-graph construction, sampling, and broader uncertainty definitions for sophisticated reasoning.
- 35.3 on BrowseComp-EN, 44.1 on BrowseComp-ZH, and 30.6 on HLE are achieved by WebSailor-V2-30B-A3B, outperforming existing open-source agents.
- WebSailor-V2-30B-A3B exceeds DeepSeek-V3.1’s 30.0 on BrowseComp-EN and 29.8 on HLE despite DeepSeek-V3.1’s 671B size.
2 Agentic Framework
WebSailor-V2 uses ReAct as a deliberately simple and universal agent framework. The agent iteratively samples thoughts and actions from prior context, interacts with tools, and stops after producing a final answer.
- ReAct is chosen as the agent foundation because its simplicity and universality expose intrinsic model capabilities and pipeline efficacy.
- A complete trajectory contains alternating thoughts, actions, and observations across T iterations.
- At each step, the policy samples the next thought and action using the entire preceding context.
- The toolkit provides search, webpage visits, Google Scholar, Python execution, and a terminal final-answer action.
3 SailorFog-QA-V2
SailorFog-QA-V2 constructs dense knowledge graphs and samples structurally diverse subgraphs to generate challenging information-seeking questions. It expands beyond acyclic structures and conventional obfuscation to cover richer reasoning patterns and uncertainty types.
- SailorFog-QA-V2 generates QA pairs from a dense knowledge graph containing real internet information.
- The knowledge graph supports sampling structurally distinct subgraphs that probe different reasoning patterns.
- Earlier seed-expansion methods tend to produce tree-like structures and struggle with cycles, feedback loops, and interdependencies.
- 3.1 Graph Construction: V2 creates denser cyclic graphs, preserves search queries and source URLs, and stores entity statistics for nuanced QA generation.
- 3.1 Graph Construction: Random-walk subgraph extraction replaces exhaustive fixed-edge enumeration because denser graphs make enumeration computationally infeasible.
- 3.3 QA Generation: QA generation distributes focus across non-isomorphic orbit nodes and introduces a wider range of defined uncertainties.
4 Agentic Post-training
The post-training pipeline combines SFT cold start, simulated and real environments, dynamic data curation, and a GRPO-derived RL algorithm. Simulation improves speed and control, while training data and environment stability are treated as central determinants of RL effectiveness.
- 4.1 SFT Cold Start: SFT first equips the agent with a robust initial policy using synthetic SailorFog-QA-V2 trajectories generated with rejection sampling.
- 4.2 Agentic Reinforcement Learning: Simulation supports rapid, low-cost, controlled experimentation because real web APIs impose costs, QPS limits, and inconsistent outputs.
- 4.2 Agentic Reinforcement Learning: Real-world training must manage instability across search sources, webpage parsers, and code execution components.
- 4.2 Agentic Reinforcement Learning: Training dynamics drive an automated synthesis-and-filtering loop that adjusts the dataset in real time and supports training stability.
- 4.3 RL Algorithm: The RL method adapts GRPO for agentic training.
- 4.3 RL Algorithm: The objective uses importance ratios and advantage estimates for policy optimization.
- 4.3 RL Algorithm: Strictly on-policy sampling, token-level policy gradients, leave-one-out variance reduction, and conservative negative-sample handling shape the training regimen.
5 Experiments
WebSailor-V2 is evaluated across challenging web-agent benchmarks against open-source and proprietary baselines, achieving strong results through SFT cold start, RL, and expanded context. Analyses show benchmark-dependent RL effects, sustained exploration, and gains from longer contexts.
- Main Results: WebSailor-V2 leads open-source agents across diverse web-agent benchmarks and surpasses proprietary systems on xbench-DeepSearch and GAIA.BrowseComp-EN and BrowseComp-ZH scores are 35.3 and 44.1, respectively.
- Main Results: 30.6 on HLE establishes a new state-of-the-art, exceeding the 671B-parameter DeepSeek-V3.1 and OpenAI-o3.The result is reported on a benchmark testing deep academic and logical reasoning.
- Main Results: SFT alone reaches 24.4 on BrowseComp-EN and 23.9 on HLE, providing the competent initial policy needed for sparse-reward RL exploration.The authors describe SFT as a prerequisite for RL stabilization and convergence on complex open-ended tasks.
- Proprietary-Agent Comparison: 48.9 on DeepResearch Bench places WebSailor-V2 second to Gemini-2.5-pro-DeepResearch at 49.7.The authors attribute the gap primarily to less emphasis on stylistic final-report quality.
- Detailed Analyses: RL reward rises with training steps, while BrowseComp-EN and BrowseComp-ZH validation performance also trends upward with oscillations.The reported pattern indicates learning within the training distribution and corresponding validation improvement.
- Detailed Analyses: On difficult BrowseComp tasks, both pass@1 and pass@3 rise, whereas simpler xbench-DeepSearch and GAIA tasks mainly improve in pass@1.The authors interpret this difference as capability expansion on difficult tasks versus sampling-efficiency gains on simpler ones.
- Detailed Analyses: Policy entropy remains high and oscillatory throughout RL, preserving exploration rather than converging prematurely to a deterministic policy.The authors connect this behavior to the stochasticity and complexity of the real-world web environment.
- Detailed Analyses: Nearly 90% of correctly solved instances finish within 64k context, and accuracy at the 32k limit is around 16 on BrowseComp-EN.Accuracy increases with context length before gradually converging; WebSailor-V2 improves over WebSailor-V1 despite using a smaller MoE model.
6 Conclusion
WebSailor-V2 combines novel data construction with refined training to build a strong open-source deep research agent. The authors argue that high-quality data and a stable training environment are more critical than the specific algorithm.
- WebSailor-V2 combines the SailorFog-QA-V2 data construction scheme with a refined training strategy.
- The agent is built on Qwen3-30B-A3B and rivals advanced proprietary Deep Research agents while surpassing previous open-source solutions.
- The authors frame high-quality agent development as a complex system-engineering challenge in which instability can produce erroneous reward signals.
- The paper argues that high-quality data and a stable training environment matter more than the specific training algorithm.
A Related Work
Recent open-source web-agent progress centers on data construction, training methods, and inference paradigms, yet a performance gap with proprietary systems has persisted. WebSailor-V2 targets this gap using a deliberately simple ReAct foundation and a dedicated data-and-training pipeline.
- Open-source web-agent research has advanced primarily through data construction, training methodologies, and inference paradigms.
- Data construction for web agents: Graph-based and easy-to-hard approaches represent two broad paradigms for constructing web-agent training data.
- Agent training strategies: A two-stage pipeline combining SFT cold start with RL policy refinement has become standard, with many implementations using GRPO variants.
- Inference paradigms: Inference choices range from vanilla ReAct to context-engineering, multi-agent, and GUI-focused paradigms.
- WebSailor-V2 addresses the persistent proprietary-versus-open-source performance gap with a moderately sized open-source model and meticulous training.
B Experimental Details
WebSailor-V2 agents operate through four web and computation tools within a ReAct-style setup. The reported training configuration uses Megatron for SFT and rLLM for RL with specified batch sizes, learning rates, and decoding parameters.
- The toolkit provides search, visit, Google Scholar, and a Python interpreter as the agent’s four primary tools.
- Search queries Google and returns the top-10 results for each query, including titles, snippets, and URLs.
- Visit retrieves web-page content through Jina and returns a summary extracted for the agent’s specified goal.
- Google Scholar supports academic information retrieval across articles, theses, books, and conference papers.
- The sandboxed Python interpreter lets the agent execute self-generated code for calculations, data analysis, and logical reasoning.
- SFT uses batch size 64 and learning rate 5e-6, while RL uses batch size 128, learning rate 1e-6, temperature 1.0, and topp = 1.0.
C Case Study
The case study shows the agent solving a complex company-identification query through decomposition, exploratory search, strategic refinement, and cross-verification. It ultimately identifies FormFactor, Inc. by matching corporate, revenue, leadership, and litigation clues.
- Case-solving process: The agent decomposed the unstructured query into clear, verifiable conditions before searching.This created manageable sub-tasks for the company-identification problem.
- Case-solving process: Broad initial searches returned irrelevant results, prompting the agent to adjust its search strategy.The exploratory queries included a generic former-employee class-action settlement search.
- Evidence verification: A precise search identified FormFactor’s founder transition to executive chairman and a new CEO appointment at the beginning of fiscal third quarter 2008.This became the breakthrough candidate match.
- Case-solving process: The complete workflow combined clue decomposition, strategic search pivots, precise retrieval, and cross-verification rather than immediate answer retrieval.The case presents this sequence as evidence of advanced research-assistant competencies.
- Evidence verification: The legal clues also matched FormFactor: a former employee filed a class action in August 2013, the court certified the class in January 2016, and the settlement required $1.5 million.The case conclusion identifies FormFactor, Inc. as the publicly traded company.
- Evidence verification: FormFactor matched the corporate and revenue clues, including Delaware incorporation in 1993, three customers contributing 47.3% of fiscal 2006 revenue, and four customers contributing 72.8% in 2005 and 64.8% in 2004.The case study uses these figures to verify the target against the query’s historical constraints.