SearchArt: Training Long-Horizon Search Agent with Scalable Synthetic and Verified Task

2026-07-25

AI Summary

This work addresses the challenges of scarce scalable high-quality data and the difficulty of evaluating intermediate reasoning behaviors in long-horizon search tasks. The authors propose a scalable framework that synthesizes large-scale, complex search task data through automated evidence graph generation and trajectory verification mechanisms. Coupled with a multi-stage post-training pipeline—comprising supervised fine-tuning and reinforcement learning—the framework trains search agents endowed with adaptive planning and iterative reasoning capabilities. Remarkably, using only a 27B-parameter model, the approach achieves scores of 74.39, 70.06, and 52.55 on BrowseComp-ZH, BrowseComp, and DeepResearch-Bench, respectively, matching or surpassing the performance of leading closed-source systems.

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.

Problem

• long-horizon search
• search agent training
• scalable task synthesis
• reasoning evaluation
• tool-use behavior

Innovation

• long-horizon search agent
• verification-driven synthesis
• evidence graph
• multi-stage training
• scalable task generation