WebSailor-V2: Bridging the Chasm to Proprietary Agents via Synthetic Data and Scalable Reinforcement Learning

2025-09-16

AI Summary

Open-source large language models (LLMs) exhibit insufficient systematic uncertainty-reduction capabilities—particularly in high-uncertainty, complex information-seeking tasks—lagging significantly behind proprietary agent systems. To address this, we propose WebSailor, a post-training framework comprising two key components: (1) the construction of challenging synthetic web-browsing scenarios that integrate structured sampling with information obfuscation to emulate real-world ambiguity; and (2) DUPO (Duplicating Sampling Policy Optimization), an efficient reinforcement learning algorithm enabling cold-start training and scalable agent adaptation. WebSailor is the first method to enable open-source agents to match the performance of proprietary systems—including DeepResearch—on benchmarks such as BrowseComp. It substantially narrows the capability gap between open- and closed-source agents and establishes a novel paradigm for uncertainty-aware modeling and reasoning in complex information retrieval.

Technology Category

• Search and Optimization: Learning to Search
• Machine Learning: Large Multimodal Models (LMMs)
• Multiagent Systems: Adversarial Agents

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
• Web Mining and Content Analysis: Large pretrained models with web data

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 open-source agents in complex information-seeking tasks, matching proprietary agents' performance and closing the capability gap.

Problem

• Bridging performance gap between open-source and proprietary AI agents
• Instilling systematic uncertainty reduction in large language models
• Enhancing complex information-seeking capabilities through scalable training

Innovation

• Synthetic data generation for uncertainty reduction
• RFT cold start and scalable reinforcement learning
• Duplicating Sampling Policy Optimization algorithm