GraphReAct: A Graph Reasoning-Acting Framework for LLMs
GraphReAct, a novel framework, empowers large language models to execute multi-step reasoning on graph-structured data by blending reasoning with actions. It features a graph-oriented action space that includes two synergistic retrieval methods: topological retrieval, which focuses on local structural dependencies, and semantic retrieval, aimed at finding non-local relevant information. This methodology tackles the difficulties of applying reasoning-acting paradigms to graph learning, where data is spread across nodes and edges and represented through topology and latent forms. The framework facilitates gradual inference, enhancing the context gathered throughout the multi-step reasoning process.
Key facts
- GraphReAct is a graph reasoning-acting framework for LLMs.
- It enables step-by-step inference over graph-structured data.
- The framework uses topological retrieval and semantic retrieval actions.
- Topological retrieval captures local structural dependencies.
- Semantic retrieval accesses non-local relevant evidence.
- The work addresses underexplored extension of reasoning-acting to graph learning.
- Graph data information is distributed across nodes and edges.
- The framework refines context during multi-step inference.
Entities
Institutions
- arXiv