G-ReAct: Graph-Guided Deep Search via Structure-State Co-Evolution
A novel reasoning framework named G-ReAct has been introduced to enhance the deep search capabilities of large language models (LLMs). Effective deep search is crucial for tackling complex open-domain tasks; however, current approaches frequently depend on linear sequential reasoning, resulting in problems such as context loss, search drift, and ineffective exploration. G-ReAct mitigates these challenges by structuring reasoning as state evolution within a fixed-topology query graph. This evolving graph state clearly monitors search advancement and directs future decisions, transforming exploratory search into graph-guided reasoning under defined constraints. The framework is applicable for both training and inference, producing high-quality deep-search trajectories. The research paper can be found on arXiv under ID 2608.01324.
Key facts
- G-ReAct is a reasoning framework for deep search in LLMs.
- It uses state evolution over a fixed-topology query graph.
- It addresses context forgetting, search drift, and inefficient exploration.
- It transforms exploratory search into graph-guided reasoning.
- It supports both training and inference.
- It generates high-quality deep-search trajectories.
- The paper is on arXiv with ID 2608.01324.
Entities
Institutions
- arXiv