# Inference Scaled GraphRAG: Improving Multi Hop Question Answering on Knowledge Graphs

**Authors:** Travis Thompson, Seung-Hwan Lim, Paul Liu, Ruoying He, Dongkuan Xu  
**arXiv:** [2506.19967](https://arxiv.org/abs/2506.19967)  
**Date:** 2025-06-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.19967v1/

## Key points

- Large Language Models (LLMs) excel in language understanding and generation but struggle with knowledge-intensive reasoning tasks requiring structured context and multi-hop information.
- Retrieval-Augmented Generation (RAG) addresses this limitation by integrating retrieved context into the generation process.
- Traditional RAG and GraphRAG methods have limitations in capturing relational structures across nodes in knowledge graphs.
- Inference-Scaled GraphRAG enhances LLM-based graph reasoning by applying inference-time compute scaling, combining sequential and parallel scaling for deeper insights and improved robustness.
- Experimental results on GRBench benchmark show significant improvement in multi-hop question answering performance compared to traditional methods, highlighting the effectiveness of inference-time scaling.
- Knowledge graphs consist of nodes representing entities connected by edges denoting relations, providing a structured framework for reasoning.
- RAG integrates external information retrieval to enhance reasoning over factual knowledge within the generation pipeline.
- Inference-time scaling involves allocating additional compute resources at test time without changing model architecture, leading to enhanced performance on complex reasoning tasks.

## Summary

Large Language Models (LLMs) have shown remarkable proficiency in language understanding and generation. However, they struggle with knowledge-intensive reasoning tasks that require access to structured context and multi-hop information. To address this limitation, Retrieval-Augmented Generation (RAG) has been introduced as a solution by incorporating retrieved context into the generation process. While traditional RAG and GraphRAG methods have made progress in this area, they often fall short in capturing relational structures across nodes in knowledge graphs. To overcome this challenge, we propose a novel framework called Inference-Scaled GraphRAG. This framework enhances LLM-based graph reasoning by applying inference-time compute scaling. Our method combines sequential scaling - where the model performs step-by-step reasoning based on previous outputs for deeper insights - with parallel scaling - where multiple responses are generated independently and aggregated using strategies like majority voting for improved robustness. In experiments conducted on the GRBench benchmark, our approach significantly improves multi-hop question answering performance compared to traditional GraphRAG and prior graph traversal baselines. These results demonstrate that inference-time scaling is a practical and architecture-agnostic solution for enhancing structured knowledge reasoning with LLMs. Additionally, our study provides background information on knowledge graphs, retrieval-augmented generation (RAG), and inference-time scaling. Knowledge graphs are defined as sets of nodes representing entities connected by edges denoting relations. RAG integrates external information retrieval into the generation pipeline to enhance reasoning over factual knowledge. Inference-time scaling involves allocating additional compute resources at test time without modifying the model architecture, enabling improved performance on complex reasoning tasks. In conclusion, by incorporating both sequential and parallel scaling into the GraphRAG framework, our proposed method enables large language models to efficiently conduct multi-hop reasoning over structured knowledge graphs in a more effective manner.