SearchArt: Training Long-Horizon Search Agent with Scalable Synthetic and Verified Task
SearchArt employs verification-driven synthetic data generation and multi-stage fine-tuning, achieving 74.39 on BrowseComp-ZH with only 27B parameters.
Key Findings
Methodology
SearchArt integrates web-enhanced knowledge graph expansion, complex relation sampling, and multi-layer verification to generate large-scale, high-quality long-horizon search datasets. The approach combines web crawling, knowledge graph construction, and difficulty ranking to produce diverse, challenging tasks. Verified trajectories undergo supervised fine-tuning followed by reinforcement learning to optimize search policies. The core algorithms include relation extraction, evidence dispersion sampling, and multi-stage validation, enabling the model to learn adaptive search strategies and complex reasoning over extended interactions. With only 27B parameters, the model surpasses many larger models on multiple benchmarks.
Key Results
- On benchmarks like BrowseComp-ZH (74.39), BrowseComp (70.06), and Deepresearch-bench (52.55), SearchArt outperforms or matches state-of-the-art closed-source models such as Gemini 3.1 Pro (85.9) and GPT 5.5 (84.4), demonstrating superior long-term reasoning and evidence integration capabilities.
- The verification-driven synthesis significantly improves intermediate reasoning reliability, leading to more stable and efficient search behaviors.
- Combining supervised fine-tuning with reinforcement learning enhances the model's ability to adapt search strategies dynamically, resulting in better evidence aggregation and reasoning depth.
Significance
This work addresses the critical bottleneck of data scarcity for training long-horizon search agents by proposing a scalable synthetic data pipeline grounded in verification. It demonstrates that a relatively small model can achieve competitive performance with larger, closed-source counterparts, paving the way for more trustworthy and autonomous AI systems capable of complex scientific and exploratory tasks. The approach offers a new paradigm for building reliable, scalable, and interpretable search agents, with broad implications for research, industry, and AI safety.
Technical Contribution
The paper introduces a novel verification-driven task synthesis framework that combines knowledge graph expansion, relation sampling, and multi-layer validation to produce high-quality training data. It innovatively integrates supervised fine-tuning with reinforcement learning, guided by a composite reward that balances outcome accuracy and process reliability. The multi-stage verification pipeline ensures intermediate reasoning steps are trustworthy, reducing noise and bias. These contributions collectively enable training of more capable, stable, and adaptive search agents within manageable computational budgets, advancing the state-of-the-art in autonomous reasoning.
Novelty
This is the first comprehensive framework that systematically combines web-enhanced knowledge expansion, complex relation sampling, and multi-stage verification for synthetic long-horizon task generation. Unlike prior work focusing solely on final answer correctness, SearchArt emphasizes the quality of intermediate reasoning and evidence aggregation. Its multi-layer verification process ensures high data fidelity, which is crucial for training reliable autonomous search agents. The integration of supervised and reinforcement learning further distinguishes it from existing methods, enabling models to learn adaptive, multi-step search strategies effectively.
Limitations
- The reliance on web data quality and coverage may limit task diversity and authenticity, especially in rapidly changing or obscure domains.
- Verification mechanisms, while robust, may still produce false positives or negatives in extremely complex or ambiguous scenarios, affecting generalization.
- Training involves high computational costs due to large-scale knowledge graph construction, sampling, and multi-stage validation, which may hinder scalability for some applications.
Future Work
Future efforts will focus on improving knowledge graph dynamism, incorporating multi-modal data, and reducing computational overhead. Developing more sophisticated, scalable verification algorithms and integrating real-time knowledge updates could further enhance model robustness. Additionally, exploring meta-learning and self-supervised techniques to reduce reliance on web data and improve generalization to unseen tasks remains a promising direction. Extending the framework to multilingual and cross-domain settings will broaden its applicability.
AI Executive Summary
In recent years, large language models (LLMs) have revolutionized natural language understanding and generation, enabling increasingly autonomous AI agents for complex information retrieval and reasoning tasks. However, training effective long-horizon search agents remains a significant challenge due to the scarcity of high-quality, diverse datasets and the difficulty of reliably evaluating intermediate reasoning steps. Traditional datasets focus on final answers, neglecting the quality of search trajectories and reasoning processes, which hampers the development of stable and adaptive strategies.
SearchArt introduces an innovative framework that addresses these issues through verification-driven synthetic data generation. By leveraging web-enhanced knowledge graph expansion, complex relation sampling, and multi-layer validation, it constructs large-scale, challenging datasets that faithfully represent real-world long-horizon search tasks. The core idea is to generate diverse, controllably difficult tasks and rigorously verify the intermediate reasoning trajectories, ensuring high data fidelity.
The training pipeline combines supervised fine-tuning with reinforcement learning, guided by a composite reward that balances answer accuracy and reasoning reliability. This approach enables the model to learn adaptive search policies, evidence aggregation, and multi-step reasoning over extended interactions. Experimental results on benchmarks like BrowseComp and Deepresearch-bench demonstrate that a 27B-parameter model trained with SearchArt surpasses many larger, closed-source models, achieving scores of 74.39, 70.06, and 52.55 respectively.
This work marks a significant advance in autonomous AI research, providing a scalable, reliable, and interpretable framework for long-horizon search and reasoning. Its implications span scientific discovery, intelligent decision-making, and AI safety, offering a practical pathway toward more capable and trustworthy AI systems. Future directions include integrating multi-modal data, dynamic knowledge updates, and reducing computational costs, aiming to push the boundaries of autonomous reasoning further.
Deep Dive
Key Concepts
验证驱动合成
通过多层验证确保生成任务和推理路径的可靠性，提升数据质量。
知识图谱扩展
利用Web信息不断扩充知识图谱，增强模型的知识覆盖和推理能力。
复杂关系采样
从知识图谱中采样结构复杂、证据分散的子图，用于生成难度适中的任务。
多阶段验证
包括QA一致性、路径可靠性和证据相关性，确保中间推理的合理性。
强化学习优化
结合奖励机制，训练模型在长链推理和搜索策略上表现更优。
Open Questions Unanswered questions from this research
- 1 如何进一步提升验证机制在极端复杂或偏离常识任务中的准确性，尤其是在多源信息冲突或噪声较多的场景下。
- 2 如何降低知识图谱构建与验证的计算成本，使其更适合大规模、多任务环境的实时应用。
- 3 在多模态信息融合和动态知识更新方面，如何设计更高效的算法以增强模型的适应性和推理深度。
Applications
Immediate Applications
科学研究助手
模型可以作为科研人员的智能助手，自动检索、整合多源信息，辅助文献分析、假设验证，提升科研效率。
智能问答系统
在企业客服、教育等场景中，提供长链、多轮的深度问答，增强用户体验和信息准确性。
Long-term Vision
自主科研探索平台
未来模型可实现自主提出研究问题、设计实验方案，推动科学发现的自动化，缩短创新周期。
Abstract
Recent advances in large language models (LLMs) have enabled search agents to autonomously tackle complex tasks across extended search and reasoning horizons. However, training effective search agents remains challenging due to the lack of scalable and long-horizon tasks, and the difficulty of evaluating and correcting intermediate reasoning and tool-use behaviors. We introduce SearchArt, a scalable framework for training long-horizon search agents through verification-driven task synthesis and a multi-stage post-training pipeline. SearchArt constructs large-scale datasets for complex search-, research- and user-oriented tasks by synthesizing diverse information-seeking QA pairs and corresponding search trajectories from web documents and automatically generated evidence graphs. To ensure the reliability of the synthesized data, we design a verification pipeline that jointly evaluates QA consistency, trajectory quality, and the relevance of retrieved evidence. The verified trajectories are subsequently used in a multi-stage training process comprising supervised fine-tuning and reinforcement learning-based policy optimization. Search agents trained with SearchArt exhibit adaptive search planning, iterative evidence aggregation, and complex reasoning over extended interaction horizons. Experimental results demonstrate that, with only (Qwen3.5-) 27B parameters, SearchArt scores 74.39 on BrowseComp-ZH, 70.06 on BrowseComp, and 52.55 on Deepresearch-bench, matching or surpassing frontier closed-source agents on both deepsearch and deepresearch benchmarks.