Source-linked AI summary
WebSailor: Navigating Super-human Reasoning for Web Agent
Kuan Li, Zhongwang Zhang, Huifeng Yin, Liwen Zhang, Litu Ou, Jialong Wu, Wenbiao Yin, Baixuan Li, Zhengwei Tao, Xinyu Wang, Weizhou Shen, Junkai Zhang, Dingchu Zhang, Xixi Wu, Yong Jiang, Ming Yan, Pengjun Xie, Fei Huang, Jingren Zhou
TL;DR
Complex web information seeking remains difficult for open-source agents because they lack reasoning strategies that systematically reduce extreme uncertainty. WebSailor addresses this with graph-synthesized, obfuscated tasks, compact cold-start supervision, and DUPO for efficient agentic RL, achieving strong performance across complex and simpler benchmarks while narrowing the proprietary gap.
Problem
Open-source agents lack the sophisticated uncertainty-reducing reasoning needed for complex web information seeking, where direct inference and existing low-uncertainty training perform poorly.
Method
WebSailor combines graph-based high-uncertainty task synthesis, information obfuscation, RFT cold start, and Duplicating Sampling Policy Optimization for agentic RL.
Results
WebSailor establishes a new state-of-the-art for open-source agents, with WebSailor-7B reaching 6.7 accuracy on BrowseComp-en versus 2.5 for WebDancer-32B and 2.8 for WebThinker-RL.
Takeaways & Limitations
WebSailor closes much of the open-source capability gap, reaching parity with Doubao on BrowseComp-zh and showing strong performance on simpler information-seeking benchmarks.
Takeaways & Limitations
Filtering trajectories below 32k tokens and limiting RL to 50 steps may constrain performance on even more complex problems.
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 opensource agents in complex information-seeking tasks, matching proprietary agents' performance and closing the capability gap.
1 Introduction
WebSailor targets the unresolved gap between proprietary and open-source agents in complex web information seeking. It combines high-uncertainty data generation, compact cold-start supervision, and efficient agentic RL to improve uncertainty-reducing reasoning.
- Open-source agents remain near-zero accuracy on BrowseComp-en because existing training emphasizes low-uncertainty or structurally clear Level 1 and 2 tasks.These datasets do not expose models to the complex Level 3 challenges that dominate the benchmark.
- WebSailor generates high-uncertainty tasks by sampling subgraphs from interconnected knowledge structures built through random walks across real-world websites.The resulting combinations of entities and relationships require reasoning about previously unseen compositions rather than simple heuristics.
- Information obfuscation increases initial ambiguity, producing tasks that can require even powerful proprietary models such as o3 to make up to 40 tool calls.Structural complexity and informational ambiguity jointly intensify the uncertainty-reduction challenge.
- A modest rejection sampling fine-tuning cold start is indispensable because complex web tasks initially provide extremely sparse RL rewards.The approach uses just over 2k high-quality examples and does not heavily rely on distillation.
- Duplicating Sampling Policy Optimization uses dynamic sampling before and during training to improve the effectiveness and efficiency of agentic RL.DUPO addresses the slow training caused by multi-turn reasoning and heavy tool use.
- WebSailor models from 3B to 72B outperform open-source agents on BrowseComp-en/zh and surpass Grok-3 and DouBao when paired with browsing.They also show promising downward compatibility on GAIA, XBench-DeepSearch, and SimpleQA.
2 Problem Definition
WebSailor formulates web information seeking as iterative ReAct interaction with external tools. Unlike fixed-path multi-hop QA, BrowseComp requires adaptive search in a vast, unstructured information space.
- The agent repeatedly produces a Thought, executes a parsable Action, and receives an environmental Observation under the ReAct framework.The action space includes final answer, search, and visit operations.
- A trajectory with T iterations is represented as alternating thoughts, actions, and observations, with each step sampled from a policy conditioned on prior history.The formal trajectory uses τ_i, a_i, and o_i for the thought, action, and observation at round i.
- Multi-hop QA usually needs only one or two ReAct rounds because its actions follow a clear path, whereas BrowseComp has no predefined solution path.BrowseComp’s unstructured search space makes brute-force exploration computationally infeasible and demands adaptive search.
3 Large-scale Training Data Synthesis for Complex Reasoning
The paper constructs complex reasoning data by classifying uncertainty levels, synthesizing graph-grounded questions, and reconstructing compact solution trajectories. The pipeline targets Level 3 tasks whose coupled entities, ambiguity, and variable topology resist predefined reasoning paths.
- Task levels: Information-seeking tasks are organized into three levels according to uncertainty and the difficulty of reducing it.Level 1 is straightforward, Level 2 follows a clear multi-hop path, and Level 3 combines high uncertainty with difficult reduction.
- Task levels: Level 3 tasks require creative exploration because their entities are coupled through complex, emergent relationships without a predefined reasoning path.These reasoning patterns are difficult to specify manually.
- Graph synthesis: The graph construction process uses fuzzy Wikidata seeds, simulated web browsing, extracted entities and relations, and stochastic expansion to create densely interconnected structures.This discourages simple linear chains and supplies the structural foundation for hard-to-reduce uncertainty.
- Question synthesis: Questions are generated from diverse sampled subgraphs and made more ambiguous by obfuscating dates, names, relationships, and quantitative attributes.The resulting tasks combine coupled entities with deliberately vague or partially masked information.
- Question synthesis: Diverse subgraph topologies produce reasoning ranging from multi-step deduction to compositional and comparative analysis.The approach is grounded in the real-world internet and supports scalable data synthesis as potential subgraphs grow non-linearly with graph size.
- Trajectory reconstruction: For cold-start supervision, the pipeline retains successful expert action-observation traces and reconstructs concise thoughts with a separate instruction-following model.The short-CoT reconstruction supplies logical justifications while avoiding verbose native reasoning outputs.
4 Reinforcement Learning with Cold Start
WebSailor trains complex web agents through a modest RFT cold start followed by RL, while DUPO improves efficiency by filtering uninformative cases and duplicating informative samples. The pipeline supervises tool-use decisions, validates trajectory format and answers, and masks environment observations from policy loss.
- A two-stage process first uses modest RFT to establish tool-use capabilities and a long-horizon reasoning skeleton, then applies RL for further refinement.
- RFT retains only correct expert trajectories, discards those exceeding 32k tokens, and filters for task complexity.
- The objective trains thoughts and actions by masking environment-observation tokens from the loss calculation.
- Agentic RL is slowed by multi-turn tool interaction and sequential replacement of uninformative rollouts in dynamic sampling.
- DUPO filters cases with all 8 rollouts correct, then duplicates same-batch samples with non-zero reward variation instead of padding.
- The reward combines rule-based format validation with LLM-judged answer validation to assess trajectory compliance and correctness.
5 Experiments
The experiments evaluate WebSailor across challenging information-seeking benchmarks and analyses of task difficulty, reinforcement learning, and training initialization. Results show strong performance on complex browsing tasks, with limitations from context length and training efficiency.
- Experimental Setup: The evaluation covers four challenging benchmarks and compares WebSailor with open-source agents, direct-inference models, and proprietary browsing agents.The reported metrics use pass@k by default, with pass@1 evaluated at non-zero temperature and accuracy judged by an LLM.
- Main Results: Direct inference performs poorly on BrowseComp-en/zh, indicating that complex information retrieval requires dynamic interaction with the web.Even strong proprietary models such as GPT-4.1 often achieve near-zero accuracy on these benchmarks.
- Main Results: 26.3 is DeepSeek-R1’s score on BrowseComp-zh, exceeding other direct-inference models in its category.The passage attributes this relative strength to improved decomposition and uncertainty reduction without external tools.
- Main Results: WebSailor-7B achieves 6.7 accuracy on BrowseComp-en, outperforming WebDancer-32B at 2.5 and WebThinker-RL at 2.8.The reported comparison supports the authors’ claim that performance gains are not merely an artifact of model scale.
- Main Results: WebSailor-72B matches Doubao on BrowseComp-zh, while DeepResearch remains ahead.This result is presented as closing the gap between open-source and proprietary systems on a difficult browsing benchmark.
6 Related Work
Related work traces information-seeking benchmarks from structured or parametric-knowledge tasks toward complex, nonlinear reasoning. It also contrasts opaque proprietary web agents with the more accessible open-source research ecosystem.
- Information-seeking benchmarks: Early benchmarks such as NQ, TriviaQA, HotpotQA, and Musique generally involve uncertainty reducible through structured queries or parametric knowledge.These datasets represent earlier stages in the evolution of information-seeking evaluation.
- Information-seeking benchmarks: GAIA and Xbench-DeepSearch extend evaluation toward multimodal, deep-search, tool-use, dynamic, and professionally annotated information-seeking tasks.The cited benchmarks target increasingly complex agent capabilities.
- Web agents: Proprietary agents show strong complex-web performance but opaque architectures and training methods, whereas open-source agents support more collaborative research.The comparison frames transparency and accessibility as distinguishing characteristics of the two communities.
7 Conclusion
WebSailor presents a full agentic post-training pipeline centered on uncertainty reduction, spanning QA construction, training-data synthesis, RFT cold start, and more efficient RL. It reports strong performance on simple and complex information-seeking benchmarks, with reasoning and tool-use capabilities described as surpassing human levels.
- Conclusion: WebSailor integrates QA construction, comprehensive training-data synthesis, RFT cold start, and improved RL efficiency into a complete agentic post-training pipeline.The pipeline is framed around uncertainty reduction in information seeking.
- Conclusion: WebSailor performs strongly on both simple and complex information-seeking benchmarks.The conclusion reports this as a cross-benchmark outcome without giving specific scores.
- Conclusion: WebSailor exhibits reasoning and tool-use capabilities that the paper describes as surpassing human levels.
- Conclusion: The authors identify more complex, higher-uncertainty tasks and more effective, efficient RL training as key directions for future agentic post-training.They also plan to extend this exploration beyond information seeking toward superhuman performance across more dimensions.
A.1 Tools
WebSailor uses search and visit as its two web-interaction tools. Search retrieves Google results for queries, while visit retrieves and summarizes targeted content from specified web pages.
- Tools: WebSailor uses two tools: search for information retrieval and visit for accessing specific web pages.
- Tools: Search accepts queries, supports multiple simultaneous searches, and returns the top 10 results per query with titles, snippets, and URLs.
- Tools: Visit takes web pages and dedicated visit goals, retrieves full page content with Jina, and uses Qwen-2.5-72B to extract relevant information.
A.2 QA Construction
The QA construction process samples subgraphs from random-walk-generated entity graphs. It begins with rare entities, expands through related entities and their features, and stops when the graph reaches a predefined edge count.
- QA Construction: Each QA is constructed by sampling a subgraph from a graph whose nodes are entities and edges are relationships.
- QA Construction: Graphs are generated by performing a random walk starting from a rare entity obtained through Wikidata’s SPARQL service.
- QA Construction: The initial entity’s features are gathered with search and visit tools, after which it becomes the expansion node.
- QA Construction: Related entities are obtained from the expansion node’s features, and their features are then collected.
- QA Construction: At each iteration, the process probabilistically chooses either a new related entity or a previously visited node as the next expansion node.
- QA Construction: Steps of obtaining and expanding entities repeat until the graph reaches a predefined number of edges.
A.3 ReAct Trajectories
WebSailor implements ReAct trajectories with Qwen-Agent and limits each trajectory to at most 30 tool calls. A trajectory alternates thinking, tool calls, and tool responses before producing an answer.
- ReAct Trajectories: The ReAct framework is implemented through Qwen-Agent, with a maximum of 30 tool calls per trajectory.
- ReAct Trajectories: A trajectory begins with a thinking block followed by a tool call.
- ReAct Trajectories: Each tool call specifies a tool name and parameterized arguments, followed by a tool response.
- ReAct Trajectories: The cycle of thinking, tool calls, and responses may repeat before the trajectory ends with an answer.
A.4 Training Details
The supplied material reports concrete SFT/RL implementation settings, DUPO’s dynamic sample handling, and a BrowseComp-en case-study answer. The case study identifies the first jointly purchased computer as an Atari 130XE.
- Training configuration: SFT uses batch size 32, learning rate 5e-6 with a 1e-10 minimum, warmup plus cosine decay, and weight decay 0.1.
- Training configuration: RL uses eight rollouts per group, temperature 1.0, topp 1.0, batch size 128, mini-batch size 32, and learning rate 1e-6.
- Case study: The BrowseComp-en case study answers that the first computer purchased jointly with the developer’s father was an Atari 130XE.The reasoning trace links this answer to identifying Joey Hess from multiple biographical clues and locating a blog entry describing the purchase.