跳转至

文章背景与核心概要

随着大语言模型（LLM）在复杂开放域任务中的广泛应用，如何提升其“深度搜索”能力成为了研究热点。传统的基于LLM的搜索方法往往依赖线性的顺序推理，这容易导致上下文遗忘、搜索漂移以及探索效率低下等问题。

为了解决这些痛点，本文介绍了全新推理框架 G-ReAct。该框架通过在固定拓扑的查询图上进行状态演化，将非结构化的文本探索转化为纪律严明、具备约束感知的推理过程。实验结果表明，G-ReAct 在 BrowseComp-ZH 和 XBench 等基准测试中取得了最先进（SOTA）的性能，且训练数据量显著少于现有方法。

G-ReAct: Graph-Guided Deep Search via Structure-State Co-Evolution

arXiv: 2608.01324
Subjects: Artificial Intelligence (cs.AI)
Authors: Shaoxiong Yang, Mengyuan Zhang, Shaojun Lin, Chao Li, Wei Liu, Kun Shao, Jian Luan
Submitted: 2 Aug 2026 (v1), 19 Aug 2026 (v3)

Summary

G-ReAct 是一种新颖的推理框架，旨在增强大语言模型（LLM）针对复杂、开放域任务的“深度搜索”能力。传统的基于LLM的搜索通常依赖于线性的、顺序的推理，这经常导致上下文遗忘、搜索漂移和低效的探索。

G-ReAct is a novel reasoning framework designed to enhance the "deep search" capabilities of Large Language Models (LLMs) for complex, open-domain tasks. Traditional LLM-based search often relies on linear, sequential reasoning, which frequently leads to context forgetting, search drift, and inefficient exploration.

G-ReAct 通过将推理组织为“固定拓扑查询图上的状态演化”来解决这些问题。通过显式地通过图结构追踪搜索进度，该模型将无结构文本探索转变为受到约束感知的规范推理。该框架既支持监督微调（使用高质量生成的轨迹），也支持无需进一步训练的推理时引导。实验结果表明，该模型在 BrowseComp-ZH 和 XBench 等基准测试上实现了最先进的性能，同时使用的训练数据显著少于现有方法。

G-ReAct addresses these issues by organizing reasoning as state evolution over a fixed-topology query graph. By explicitly tracking search progress through a graph structure, the model transforms unstructured textual exploration into disciplined, constraint-aware reasoning. The framework supports both supervised fine-tuning (using high-quality generated trajectories) and inference-time guidance without requiring further training. Experimental results show that the model achieves state-of-the-art performance on benchmarks like BrowseComp-ZH and XBench while using significantly less training data than existing methods.

Key Features

  • 图引导推理： 从基于历史的线性搜索转向结构化的状态演化。
  • 约束保持： 在长视距、多跳搜索的过程中，利用查询图来维持中间状态和约束。
  • 训练与推理的灵活性： 提供了一种双模式方法，可通过微调提升性能，亦可作为现有LLM的推理时插件。
  • 高效率： 仅使用1.9K生成的轨迹便取得了卓越的成果，优于更大的模型和经强化学习增强的基线。
  • Graph-Guided Reasoning: Shifts from linear history-based search to structured state evolution.
  • Constraint Preservation: Uses the query graph to maintain intermediate states and constraints throughout long-horizon, multi-hop searches.
  • Training & Inference Flexibility: Provides a dual-mode approach that improves performance via fine-tuning or as an inference-time plugin for existing LLMs.
  • Efficiency: Achieves superior results with only 1.9K generated trajectories, outperforming larger models and RL-enhanced baselines.

Performance Highlights

  • Qwen3-30B-A3B-Thinking-2507 的性能表现：
    • 在 BrowseComp-ZH 上准确率为 52.6%。
    • 在 XBench 上准确率为 79.0%。
  • 当在推理阶段应用时，可持续提升现有强大LLM的深度搜索能力。
  • Qwen3-30B-A3B-Thinking-2507 performance:
    • 52.6% accuracy on BrowseComp-ZH.
    • 79.0% accuracy on XBench.
  • Consistently improves the deep-search capabilities of existing strong LLMs when applied at inference time.

Access & Resources

注：本内容基于 arXiv 提交编号 2608.01324。更多详情，请参考官方 arXiv 页面。

Note: This content is based on the arXiv submission 2608.01324. For further details, please refer to the official arXiv page.