Research map / Large language models / Retrieval and knowledge

Can Knowledge-Graph-based Retrieval Augmented Generation Really Retrieve What You Need?

NeurIPS 2025 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GraphFlow, a GFlowNet-based detailed-balance method with local exploration that trains an LLM retrieval policy and flow estimator from outcome rewards to retrieve accurate and diverse documents from text-rich KGs on STaRK.

Problem

KG-based RAG struggles to retrieve accurate and diverse information for complex queries over text-rich KGs, and PRM-style step-level supervision for retrieval is expensive to obtain.

Approach

Models agent retrieval on KGs as a multi-step flow process; a flow estimator factorizes the outcome reward into intermediate state flows, and the policy is trained with a detailed-balance objective plus local exploration so retrieval samples trajectories in proportion to reward, without process-level labels.

Key results

• On STaRK-PRIME with rerank, GraphFlow reaches Hit@1 51.39 and MRR 61.37 versus ToG+GPT4o's 53.33 and 57.78, and on MAG reaches MRR 47.82 vs 39.65.
• GraphFlow achieves the highest R@20 and D-R@20 on all three datasets, e.g. 79.71 R@20 on STaRK-PRIME vs 54.35 for ToG+GPT4o.
• The paper reports that GraphFlow beats GPT-4o-based baselines by about 10% average on accuracy and diversity metrics.

Measured effects reported in the paper

Surprising findings

• Without rerank, GraphFlow has lower Hit@1 than ToG+GPT4o on STaRK-AMAZON (19.63 vs 20.67) and PRIME-Hit@1 is lower with rerank than ToG+GPT4o.
• Reported PRM baseline underperforms SFT on several metrics, e.g. STaRK-PRIME with rerank MRR 26.94 vs SFT 35.98.

Limitations

• Generalization evaluated on only two new domains.
• Baselines required hybrid ego-graph settings (2-hop) because full-KG retrieval was infeasible on STaRK.
• No error bars or statistical significance reported.

Experimental setup

Datasets: STaRK-AMAZON STaRK-MAG STaRK-PRIME

Benchmarks: STaRK-AMAZON STaRK-MAG STaRK-PRIME

Metrics: Hit@1 Hit@5 MRR R@20 D-R@20 Step-Seper Answer-Seper

Baselines: DenseRetriever G-Retriever SubgraphRAG ToG+LLaMA3 ToG+GPT4o SFT PRM

Builds on: GFlowNet Process Reward Models ToG SubgraphRAG G-Retriever

Closest papers by meaning

• A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 2025 · similarity 0.92)
• Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.92)
• GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation (NeurIPS 2025 · similarity 0.91)
• Think-on-Graph 2.0: Deep and Faithful Large Language Model Reasoning with Knowledge-guided Retrieval Augmented Generation (ICLR 2025 · similarity 0.91)
• HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation (NeurIPS 2025 · similarity 0.89)
• Adapting to Non-Stationary Environments: Multi-Armed Bandit Enhanced Retrieval-Augmented Generation on Knowledge Graphs (AAAI 2025 · similarity 0.89)