Intergrating LLM and GNN to enhance natural language understanding

Summarise with VIRAT
Authors: Ziming Wang
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Retrieval-Augmented Generation (RAG) systems have evolved from static vector retrieval to dynamic agentic workflows to handle complex, multi-hop reasoning tasks. However, existing state-of-the-art frameworks mainly adopt a single unit architecture, employing a single large-scale LLM to perform both graph navigation and answer synthesis. This identifies a capacity-task mismatch, arguing that the information filtering does not require the same parametric capacity as semantic task of reasoning. To address this, this research proposes Decoupled Navigation and Reasoning Graph Agent (DNR-Graph), a framework that decouples the agentic workflow. The architecture comprises two specialized roles: a Navigator based on a SLM and a Reasoner based on a frozen LLM. The Navigator is optimized via GRPO using rule-based rewards. This training process aligns the small model’s policy with topological constraints, enabling it to actively prune noise and construct a concise evidence chain. The Reasoner synthesizes the final answer using only high quality context selected by the Navigator. Experiments are conducted on two benchmarks: HotpotQA and 2WikiMultiHopQA. Results demonstrate DNR-Graph achieves a comparable reasoning accuracy to SOTA agents. The GRPO-trained Navigator acts as a high precision filter. These findings validate that small models by RL can effectively replace large generalist agents for graph navigation tasks, offering a computationally efficient solution for reasoning systems.

Fields of Study & Subjects

Topic Modeling Information Retrieval and Search Behavior Multimodal Machine Learning Applications

Publication details

Journal / source
DR-NTU (Nanyang Technological University)
Publication date
1 January 2026
Language
EN
Document type
Thesis