文章背景与核心概要

在大语言模型（LLM）处理需要深度搜索的开放域复杂任务时，传统的线性顺序推理往往会导致上下文遗忘、搜索漂移以及在长视野探索中的低效。为了克服这些局限性，研究人员推出了 G-ReAct，这是一种全新的推理框架，它将问题求解构建为“固定拓扑查询图上的状态演化”。

通过在演化的图状态中显式追踪搜索进展，G-ReAct 将原始文本历史探索转化为受清晰约束支配的图引导推理。该框架同时支持使用高质量轨迹进行监督微调（SFT）以及无需额外训练的推理期引导。值得注意的是，仅使用 1.9K 条轨迹对 Qwen3-30B-A3B-Thinking-2507 进行微调，就在 BrowseComp-ZH 上取得了 52.6% 的惊人准确率，在 XBench 上取得了 79.0% 的准确率，表现显著优于体量更大的开源方法。

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

📌 Summary

Large Language Models (LLMs) often struggle with open-domain complex tasks requiring deep search, primarily because linear sequential reasoning leads to context forgetting, search drift, and inefficient exploration over long horizons. To overcome these limitations, researchers introduce G-ReAct, a novel reasoning framework that structures problem-solving as state evolution over a fixed-topology query graph.

By tracking search progress explicitly through an evolving graph state, G-ReAct transforms raw text-history exploration into graph-guided reasoning governed by clear constraints. It supports both supervised fine-tuning (SFT) using high-quality trajectories and zero-extra-training inference guidance. Notably, fine-tuning Qwen3-30B-A3B-Thinking-2507 on just 1.9K trajectories yields an impressive 52.6% accuracy on BrowseComp-ZH and 79.0% on XBench, outperforming significantly larger open-source methods.

📄 Metadata

🔍 Abstract

深度搜索已成为大语言模型（LLM）解决开放域复杂任务的一项基本能力。然而，现有方法在轨迹生成和推理阶段通常都依赖于线性顺序推理，这使得在长视野、多跳搜索过程中很难持续保留中间状态和约束。因此，它们经常遭遇上下文遗忘、搜索漂移和探索效率低下的困扰。

Deep search has become a fundamental capability of large language models (LLMs) for solving open-domain complex tasks. However, existing approaches typically rely on linear sequential reasoning for both trajectory generation and inference, making it difficult to consistently preserve intermediate states and constraints throughout long-horizon multi-hop search. Consequently, they often suffer from context forgetting, search drift, and inefficient exploration.

为了解决这些局限性，我们提出了 G-ReAct，这是一种用于深度搜索的推理框架，它将推理组织为固定拓扑查询图上的状态演化。演化中的图状态显式地追踪搜索进展并指导后续决策，从而将由文本历史驱动的探索性搜索转化为在明确约束下的图引导推理。G-ReAct 同时支持训练和推理：它为监督微调生成高质量的深度搜索轨迹，并在无需额外微调的情况下为推理期搜索提供结构化引导。

To address these limitations, we propose G-ReAct, a reasoning framework for deep search that organizes reasoning as state evolution over a fixed-topology query graph. The evolving graph state explicitly tracks search progress and guides subsequent decisions, transforming exploratory search driven by textual history into graph-guided reasoning under explicit constraints. G-ReAct supports both training and inference: it generates high-quality deep-search trajectories for supervised fine-tuning and provides structured guidance for inference-time search without additional fine-tuning.

实验表明，仅使用 1.9K 条生成的微调轨迹，Qwen3-30B-A3B-Thinking-2507 在 BrowseComp-ZH 上就达到了 52.6% 的准确率，在 XBench 上达到了 79.0% 的准确率，性能超越了使用显著更大数据集（包括强化学习增强方法）训练的同类开源方法。此外，当在推理阶段应用时，G-ReAct 能够持续提升现有强大 LLM 在深度搜索任务上的性能。我们将公开所有的代码和模型权重。

Experiments demonstrate that with only 1.9K generated trajectories for fine-tuning, Qwen3-30B-A3B-Thinking-2507 achieves 52.6% accuracy on BrowseComp-ZH and 79.0% on XBench, outperforming comparable open-source methods trained on substantially larger datasets, including RL-enhanced methods. Furthermore, when applied at inference time, G-ReAct consistently improves the performance of existing strong LLMs on deep-search tasks. We will publicly release all code and model weights.

🚀 Key Innovations & Features

• 图引导推理： 从容易出错的线性文本轨迹转变为使用固定拓扑查询图的结构化方法。
• 结构-State 协同演化： 显式追踪进展并保留中间目标/约束，以克服上下文遗忘和搜索漂移。
• 双模能力：
• 监督微调（SFT）： 从高质量的生成深度搜索轨迹中学习。
• 推理期引导： 在运行过程中即时优化推理，无需额外的模型训练。
• 高数据效率： 仅使用精简的 1.9K 训练轨迹集，即超越了海量数据集和强化学习（RL）基线。

• Graph-Guided Reasoning: Shifts away from error-prone linear text trajectories to a structured approach using a fixed-topology query graph.
• Structure-State Co-Evolution: Explicitly tracks progress and retains intermediate goals/constraints to combat context forgetting and search drift.
• Dual-Mode Capability:
• Supervised Fine-Tuning (SFT): Learns from high-quality, generated deep-search trajectories.
• Inference-Time Guidance: Optimizes reasoning on-the-fly without requiring additional model training.
• High Data Efficiency: Outperforms massive datasets and reinforcement learning (RL) baselines using a compact set of only 1.9K training trajectories.