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

2026-08-02

AI Summary

Current large language models are constrained by linear reasoning in long-range multi-hop tasks, often suffering from context forgetting, search drift, and inefficient exploration. This work proposes G-ReAct, a novel framework that formulates reasoning as a state evolution process over a fixed-topology query graph, enabling co-evolution of structure and reasoning through explicit maintenance of graph-structured states. The approach provides structured guidance during inference without requiring additional fine-tuning and efficiently generates high-quality training trajectories. Using only 1.9K fine-tuning samples, Qwen3-30B-A3B-Thinking-2507 achieves 52.6% and 79.0% accuracy on BrowseComp-ZH and XBench, respectively—significantly outperforming open-source methods that rely on substantially larger training sets—and demonstrates consistent performance gains even on strong base models.

Abstract

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. To address these limitations, we propose $\textbf{G-ReAct}$, a reasoning framework for deep search that organizes reasoning as $\textbf{state evolution over a fixed-topology query graph}$. 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. 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.

Problem

• deep search
• reasoning
• state preservation
• search drift
• multi-hop reasoning

Innovation

• graph-guided reasoning
• state evolution
• deep search
• structured constraints
• query graph