Organizations: School of Automation and Intelligent Sensing, Shanghai Jiao Tong University, Shanghai 200240, China. · Department of Data Science and AI, Monash University, Melbourne, Australia. · Tencent Youtu Lab, Shanghai 200240, China. · School of Computer and Data Science, Fuzhou University, Fuzhou 350116, China. · School of Information and Communication Technology, Griffith University, Queensland, Australia.
Large Language Models (LLMs) have shown remarkable capabilities across tasks, but they still struggle with complex reasoning that requires integrating scattered knowledge. Graph Retrieval-Augmented Generation (GraphRAG) addresses this challenge by organizing fragment knowledge as graphs, while recent Graph Foundation Models (GFMs) further improve GraphRAG with graph-native retrieval and reasoning capabilities. However, existing GraphRAG methods usually adopt decoupled retrieve-then-generate pipelines, which create objective misalignment between retrieval and reasoning, and training GFM retrievers typically depends on expensive human annotations. To address these limitations, we propose \ourmethod, a unified end-to-end GraphRAG framework that tightly couples a GFM retriever with an LLM generator. Given a query, the GFM retriever first reasons over the graph to score and retrieve relevant document nodes. These node embeddings are then passed through a fully differentiable graph bridge, which projects them into soft-weighted graph prefixes according to retrieval scores and combines them with raw document texts to form graph-augmented prompts. In this way, structural information is injected directly into LLM generation, while the final answer loss can be back-propagated through the bridge to jointly optimize retrieval and reasoning. To stabilize training and reduce reliance on costly annotations, we further introduce a semantic-guided distillation module that provides auxiliary supervision using pseudo-labels derived from query-document semantic similarities. Extensive experiments on three multi-hop QA datasets and three out-of-domain GraphRAG benchmarks show that \ourmethod consistently outperforms strong baseline methods and generalizes well, which demonstrates the effectiveness of end-to-end GraphRAG with tightly integrated GFM retrieval and LLM reasoning.
Figures & tables
Fig. 1 : Comparison between the current two-stage GraphRAG and the proposed UniGFM-RAG . Existing GraphRAG methods usually employ a decoupled retrieve-then-generate pipeline, which often results in objective misalignment. In contrast, UniGFM-RAG introduces a unified end-to-end GraphRAG framework that tightly couples a Graph Foundation Model (GFM) retriever with a Large Language Model (LLM) generator.
Notation
Description
Knowledge Graph & Query
G
The heterogeneous knowledge graph
v∈V
A generic node v in the node set V
r∈R
A relation type r in the relation set R
(v,r,v′)∈E
An edge (v,r,v′) in the edge set E
tv∈T
The node type of v , belonging to the type set T
TABLE I : Glossary of commonly used notations.
Fig. 2 : The overall framework of UniGFM-RAG , which comprises of three core components: (a) a graph foundation model retriever , which reasons over the structural knowledge for accurate document retrieval; (b) a graph bridge module , which seamlessly connects the retriever and the generator into a unified framework; (c) a graph-conditioned LLM generator , which incorporates the structural information to guide the final answer generation. In addition, we introduce a semantic-guided distillation module to provide auxiliary supervision for training GFMs.
Dataset
# Query
# Document
HotpotQA [ 53 ]
1,000
9,221
MuSiQue [ 54 ]
1,000
6,119
2Wiki [ 55 ]
1,000
11,656
G-bench (Novel) [ 56 ]
2,010
461
G-bench (Medical) [ 56 ]
2,062
2,406
G-bench (CS) [ 57 ]
1,018
24,534
TABLE II : Statistics of the evaluation datasets.
Method
HotpotQA
MuSiQue
2Wiki
EM
F1
EM
F1
EM
F1
Pure LLM Methods
GPT-4o-mini (ZS) [ 58 ]
28.6
41.0
11.2
36.3
30.2
36.3
Qwen3-8B (ZS) [ 2 ]
15.1
23.7
1.3
8.1
23.2
27.9
Qwen3-8B (FT) [ 2 ]
19.1
27.0
2.7
10.7
25.8
29.0
Traditional RAG Methods
TABLE III : QA reasoning performance comparison on multi-hop datasets.
HotpotQA
MuSiQue
2Wiki
Method
R@2
R@5
R@2
R@5
R@2
R@5
Traditional RAG Methods
BM25 [ 59 ]
55.4
72.2
32.3
41.2
51.8
61.9
ColBERTv2 [ 60 ]
64.7
79.3
37.9
49.2
59.2
68.2
Qwen3-Emb [ 49 ]
67.5
80.0
41.9
55.7
63.6
70.0
Graph-enhanced Methods
TABLE IV : Retrieval performance comparison on multi-hop QA datasets. Recall@ k (R@ k ) is used as the evaluation metric.
Method
G-bench (ACC)
G-bench (Recall)
Novel
Medical
CS
Novel
Medical
Pure LLM Methods
GPT-4o-mini (ZS) [ 58 ]
51.4
67.1
70.7
-
-
Qwen3-8B (ZS) [ 2 ]
31.2
38.5
63.5
-
-
Qwen3-8B (FT) [ 2 ]
42.2
39.4
24.9
-
-
Traditional RAG Methods
TABLE V : Performance comparison on the G-bench dataset [ 56 , 57 ] . QA reasoning is measured by Accuracy (ACC), and retrieval is measured by Recall.
Fig. 3 : Ablation studies of UniGFM-RAG .
Fig. 4 : Parameter analysis on hyperparameter on k and λ .
Variant
HotpotQA
MuSiQue
2Wiki
EM
F1
EM
F1
EM
F1
GFM-RAG
46.3
59.5
23.5
34.8
60.0
66.8
UniGFM-RAG
Qwen3-8B
63.4
75.7
31.1
44.0
61.7
67.1
Qwen3-4B
61.2
73.2
26.7
37.5
59.3
64.67
Llama-3.1-8B
59.5
72.3
29.4
39.2
59.0
63.5
TABLE VI : Different backbone models in UniGFM-RAG .
Fig. 5 : Efficiency and performance comparison on G-bench (CS).
Question
What American professional Hawaiian surfer born 18 October 1992 won the Rip Curl Pro Portugal?
Gold Answer
John John Florence
Supporting Docs.
[“MEO Rip Curl Pro Portugal”, “John John Florence”]
GFM-RAG (Baseline Method)
Retrieved Docs.
[“MEO Rip Curl Pro Portugal”, “Reef Hawaiian Pro” ] (Recall@2=0.5)
Predicted Answer
document does not contain information about John John Florence ✗ Incorrect
UniGFM-RAG (Ours)
TABLE VII : Case study comparing the retrieval and reasoning performance of UniGFM-RAG with GFM-RAG on the HotpotQA dataset.
Question
Which season did Brian Billick coach the Baltimore Ravens though in 2001?
TABLE IX : Statistics of the query-doc pairs and KGs used for training.
Fig. 6 : The graph-augmented prompt for LLM Reasoning.
Fig. 7 : Fine-grained comparison of UniGFM-RAG and GFM-RAG under different reasoning depths. EM and F1 scores are reported for 2-hop, 3-hop, and 4-hop questions across datasets.
Question
What UK label was bought by the broadcast company that, along with ABC and the original network of Undercovers, is one of the major broadcasters based in New York?
Gold Answer
Oriole Records
Supporting Docs.
[“New York City_5”, “Undercovers (TV series)”, “Sony Music_4”]
GFM-RAG (Baseline Method)
Retrieved Docs.
[ “Undercovers (TV series)” , “Neil Reid”, “999 ABC Broken Hill”, “Julie (TV series)”, “The Night Shift (season 3)”] (recall@5=1/3)
Predicted Answer
ITV ✗ Incorrect
UniGFM-RAG (Ours)
TABLE X : Case study comparing the retrieval and reasoning performance of UniGFM-RAG with GFM-RAG on the MuSiQue dataset ( 3-hop ).
Question
A country’s military branch, the equivalent of which in the US helped the Old West’s local police, was unprepared for the invasion of the country that developed the ZH-29. The unprepared country was the only communist country to have an embassy where?
School of Computer Science, Nanjing University of Information Science and Technology, Nanjing 210044, China · State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210023, China · National Institute of Healthcare Data Science, Nanjing University, Nanjing 210093, China +2