WebSailor: Navigating Super-human Reasoning for Web Agent

2025-07-03

AI Summary

Open-source large language models (LLMs) struggle with extreme uncertainty in complex web browsing tasks, particularly in information-seeking scenarios requiring multi-step reasoning and dynamic environment interaction. Method: We propose a systematic uncertainty-mitigation agent reasoning paradigm. Our approach constructs high-uncertainty task environments and introduces an end-to-end post-training framework integrating structured sampling, information obfuscation, RFT cold-start initialization, and a novel proxy reinforcement learning algorithm—DUPO. Contributions/Results: (1) We design the first scalable reasoning architecture explicitly targeting extreme uncertainty; (2) we establish the first post-training paradigm enabling efficient, robust training of web agents; and (3) we bridge the critical capability gap between open- and closed-source agents on complex retrieval benchmarks. Empirically, our method achieves performance on par with proprietary systems (e.g., DeepResearch) on BrowseComp, significantly outperforming prior open-source agents.

Technology Category

• Reasoning under Uncertainty: Uncertainty Representations
• Multiagent Systems: Multiagent Systems under Uncertainty
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

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.

Problem

• Overcoming human cognitive limits in LLM training
• Reducing uncertainty in vast information navigation
• Bridging performance gap between open-source and proprietary agents

Innovation

• Generates high-uncertainty tasks via structured sampling
• Uses RFT cold start for initial model training
• Applies Duplicating Sampling Policy Optimization (DUPO)