Large Language Models (LLMs) often struggle with inherent knowledge boundaries and hallucinations, limiting their reliability in knowledge-intensive tasks. While Retrieval-Augmented Generation (RAG) mitigates these issues, it frequently overlooks structural interdependencies essential for multi-hop reasoning. Graph-based RAG approaches attempt to bridge this gap, yet they typically face trade-offs between accuracy and efficiency due to challenges such as costly graph traversals and semantic noise in LLM-generated summaries. In this paper, we propose HyperNode Expansion and Logical Path-Guided Evidence Localization strategies for GraphRAG (HELP), a novel framework designed to balance accuracy with practical efficiency through two core strategies: 1) HyperNode Expansion, which iteratively chains knowledge triplets into coherent reasoning paths abstracted as HyperNodes to capture complex structural dependencies and ensure retrieval accuracy; and 2) Logical Path-Guided Evidence Localization, which leverages precomputed graph-text correlations to map these paths directly to the corpus for superior efficiency. HELP avoids expensive random walks and semantic distortion, preserving knowledge integrity while drastically reducing retrieval latency. Extensive experiments demonstrate that HELP achieves competitive performance across multiple simple and multi-hop QA benchmarks and up to a 28.8× speedup over leading Graph-based RAG baselines.
Figures & tables
Figure 1: Overview of the HELP framework. The workflow consists of three stages: (I) Knowledge Graph Construction, utilizing OpenIE to build the Triple-to-Passage Index; (II) Iterative HyperNode Expansion, which iteratively chains triples as HyperNodes while pruning irrelevant ones to maintain reasoning coherence; and (III) Logical Path-Guided Evidence Localization, where HyperNodes are grounded back to the original passages for precise evidence retrieval via a hybrid mechanism.
Simple QA
Multi-Hop QA
Retrieval
NQ
PopQA
MuSiQue
2Wiki
HotpotQA
LV-Eval
Avg
Simple Baselines
None
54.9
32.5
26.1
42.8
47.3
6.0
34.9
Contriever ( Izacard et al., 2021 )
58.9
53.1
31.3
41.9
62.3
8.1
42.6
BM25 ( Robertson and Walker, 1994 )
59.0
49.9
28.8
51.2
63.4
5.9
43.0
GTR ( Ni et al., 2022 )
59.9
56.2
34.6
52.8
62.8
7.1
45.6
Table 1: Performance (F1 scores) comparison of various retrieval methods across different tasks, with all models evaluated using Llama-3.3-70B-Instruct as the backbone. Results are directly reported from HippoRAG2 ( Gutiérrez et al., 2025 ) by default, while methods marked with ( ∗ ) denote our local reproduction. The best results are in bold , and the second best are underlined . The † symbol indicates a failure in graph construction due to large corpus size.
Figure 2: Retrieval efficiency on PopQA (Simple QA) and 2Wiki (Multi-Hop QA). Absolute retrieval time (in seconds) for processing 1,000 queries are annotated above the bars, highlighting that HELP reduces the total retrieval latency to under 90 seconds.
Quota(M)
EM(%)
F1(%)
Recall@5(%)
0
57.5
61.55
76.25
1
57.7
62.19
76.52
2
63.0
69.52
84.95
3
64.3
71.01
89.00
4
66.6
73.90
92.15
5
65.9
73.09
91.65
Table 2: Ablation study of the Hybrid Retrieval Strategy on the 2Wiki dataset. M denotes the primary quota of passages assigned to the logical path-guided retrieval, while the total context size is fixed at 5. M=0 corresponds to a purely dense retrieval baseline.
Figure 3: Experimental analysis of expansion hops N on HotpotQA dataset. The bars indicate retrieval time (left axis), while the line tracks the QA F1 score (right axis).
Figure 4: Hyperparameter sensitivity analysis of the Initial Triple Size ( n ) and Hypernode Beam Size ( k ) on MuSiQue dataset. The heatmap reports the F1 scores (%), demonstrating the impact of varying the seed set size and pruning threshold.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Description
Value
N
Number of expansion hops
2
n
Initial seed triple size
3
k
Hypernode beam search size
50
M
Quota for logical path-guided passages
4
L
Total context size ( M + Dense passages)
5
Appendix
Table 3: Hyperparameter settings for the HELP framework.
Simple QA
Multi-Hop QA
Retrieval
NQ
PopQA
MuSiQue
2Wiki
HotpotQA
LV-Eval
Avg
Graph-based RAG Methods
HippoRAG2 ( Gutiérrez et al., 2025 )
40.4 / 54.4
42.2 / 56.1
34.3 / 45.8
60.2 / 69.4
58.4 / 72.9
7.26 / 9.67
40.5 / 51.4
LinearRAG ( Zhuang et al., 2025 )
31.6 / 43.8
6.6 / 21.9
18.7 / 28.1
42.4 / 53.2
39.6 / 52.4
6.5/ 10.7
24.2 / 35.0
HyperGraphRAG ( Luo et al., 2025 )
30.3 / 42.1
7.9 / 22.5
22.7 / 36.9
45.5 / 58.6
54.3 / 70.2
N/A †
-
HELP (Ours)
43.0 / 56.9
44.5 / 57.4
33.7 / 44.7
62.0 / 69.5
60.4 / 73.9
10.5 / 12.9
42.4 / 52.6
Appendix
Table 4: Performance comparison (EM / F1 scores) of various retrieval methods across different tasks, with all models evaluated using Qwen3-30B-A3B-Instruct-2507 as the backbone. In each cell, the first value represents the Exact Match (EM) score, and the second represents the F1 score. The best results are in bold . The † symbol indicates a failure in graph construction due to large corpus size.
Question
Who is the husband of Princess Elene Of Georgia?
Ground Truth
Prince Archil of Imereti
Reasoning Chain
Princess Elene of Georgiamother ofSolomon II of Imeretihad fatherPrince Archil of Imereti
LinearRAG
Retrieved Passages: 1) ✗ Grand Duchess Elena Vladimirovna of Russia: …husband was Prince Nicholas… 2) ✗ Prince Adarnase of Kartli: …natural son of Levan of Kartli by a concubine… 3) ✗ Princess Charlotte of Württemberg: …the wife of Grand Duke Michael Pavlovich of Russia… 4) ✗ Clous van Mechelen: …is a Dutch musician, arranger, and actor… 5) ✗ Bernhard III, Prince of Anhalt-Bernburg: …was the eldest son of bernhard ii… Prediction: ✗ There is no direct information provided about the husband of Princess Elene of Georgia in the given passages.
HyperGraphRAG
Retrieved Passages: 1) ✗ Gundobad: …He was the husband of Caretene. 2) ✗ Eunoë (wife of Bogudes): Eunoë was the wife of Bogudes, King of Mauretania. 3) ✗ Engelbert III of the Mark: …Adolph was the eldest son of Count Adolph II… 4) ✗ Megingoz of Guelders: …He married Gerberga of Lorraine… 5) ✗ Princess Rodam of Kartli: …Princess Rodam married King George VII… Prediction: ✗ King George VII of Imereti
HippoRAG2
Retrieved Passages: 1) ✓ Princess Elene of Georgia: …She was the mother of Solomon II of Imereti… 2) ✗ Grand Duchess Elena Vladimirovna of Russia: …husband was Prince Nicholas… 3) ✗ Princess Rodam of Kartli: …Princess Rodam married King George VII… 4) ✓ Solomon II of Imereti: …He was born as David to Prince Archil of Imereti… 5) ✗ Grand Duchess Elena Pavlovna of Russia: …his second wife Sophie Dorothea of Württemberg… Prediction: ✗ None mentioned.
HELP
Expansion from Initial Triples as Seed HyperNodes ("princess elene of georgia", "daughter of", "heraclius ii") ("princess elene of georgia", "mother of", "solomon ii of imereti") ("princess elene of georgia", "born in", "georgia") Retrieved Passages: 1) ✓ Princess Elene of Georgia: …She was the mother of Solomon II of Imereti… 2) ✓ Solomon II of Imereti: …He was born as David to Prince Archil of Imereti… 3) ✗ Grand Duchess Elena Vladimirovna of Russia: …husband was Prince Nicholas… 4) ✗ Princess Rodam of Kartli: …Princess Rodam married King George VII… 5) ✗ Grand Duchess Elena Pavlovna of Russia: …his second wife Sophie Dorothea of Württemberg… Prediction: ✓ Prince Archil of Imereti
Appendix
Table 5: A case study illustrating how different methods handle a 2-hop query requiring cross-passage inference. HELP successfully leverages seed HyperNodes to anchor the correct reasoning path, whereas other baselines suffer from semantic drift or hallucination.