Multi-hop question answering requires retrieval agents to construct a connected evidence path across reasoning turns rather than retrieve isolated relevant evidence for a single query. However, in existing GraphRAG methods, retrieval and reinforcement learning are often optimized with mismatched signals: retrieval is driven by local semantic relevance, while outcome-level reinforcement learning assigns the reward by assessing the final answer without identifying which retrieval actions progress the reasoning path. We introduce HyperGraphPro, a progress-aware hypergraph RAG framework that focuses on evidence-path progress as the shared objective for retrieval and policy optimization. For retrieval, HyperGraphPro reranks candidate hyperedges using query-conditioned structural distinctiveness, combining entity-query relevance with a local-vs-global hypergraph incidence signal to favor entities that are informative in the current graph neighborhood. For policy optimization, HyperGraphPro augments uniform trajectory-level reward assignment with stepwise progress-aware policy optimization, which assigns step-level rewards according to the coverage of answer-guided target hyperedges. Experiments on multi-hop question answering benchmarks demonstrate that HyperGraphPro consistently improves reasoning accuracy over existing GraphRAG methods.
Figures & tables
Method
2Wiki.
HotpotQA
MuSiQue
Multihop Avg.
NQ
All Avg.
EM
F1
EM
F1
EM
F1
EM
F1
EM
F1
EM
F1
Qwen2.5-3B-Instruct
NaiveGeneration
2.34
7.59
6.25
11.16
0.78
3.67
3.12
7.47
2.34
8.90
2.93
7.83
StandardRAG
3.91
12.52
7.03
15.41
0.00
2.92
3.65
10.28
0.00
10.69
2.74
10.39
SFT
7.03
12.40
10.94
16.48
1.56
5.04
6.51
11.31
3.12
11.23
5.66
11.29
R1
20.31
28.45
20.31
25.33
3.12
8.07
14.58
20.62
11.72
21.51
13.87
20.84
Table 1: Experimental results on multihop question-answering benchmark datasets such as 2WikiMultiHopQa, HotpotQA, and MuSiQue, and a singlehop question-answering dataset such as NQ. Multihop Avg. indicates the mean performance over multihop datasets. All Avg. indicates the mean performance over all datasets.
2Wiki.
MuSiQue
Avg.
2Wiki.
MuSiQue
Avg.
Method
EM
F1
EM
F1
EM
F1
EM
F1
EM
F1
EM
F1
Qwen3-0.6B
Qwen3-1.7B
Search-R1
33.59
37.92
9.38
15.55
21.49
26.74
31.25
36.79
10.94
17.23
21.10
27.01
Graph-R1
35.94
43.66
15.62
22.21
25.78
32.94
49.22
57.49
23.44
32.88
36.33
45.19
HyperGraphPro (Ours)
43.75
51.43
26.56
33.91
35.16
42.67
54.69
62.50
32.81
41.91
43.75
52.21
Table 2: Experimental results using Qwen3-0.6B and Qwen3-1.7B. Avg. denotes the mean over 2Wiki. and MuSiQue.
2Wiki.
MuSiQue
Str. guided ret.
Progress-aware PO
EM
F1
EM
F1
50.00
57.56
32.81
40.51
✓
59.38
66.04
39.06
49.91
✓
57.81
65.88
37.50
46.22
✓
✓
63.28
69.76
42.19
52.26
Table 3: Ablation studies of our HyperGraphPro on multi-hop question answering datasets with Qwen-2.5-3B. (Str. guided ret.: Structure-guided hypergraph retrieval, Progress-aware PO: Progress-aware Step-level policy optimization).
2Wiki.
MuSiQue
Retrieval method
EM
F1
EM
F1
No Retrieval
20.31
28.45
3.12
8.07
Knowledge corpus (Ctxt.)
31.25
38.04
3.91
7.65
Hyperedge (Ctxt.)
50.78
58.13
32.03
40.31
Graph-R1 (Ctxt.)
50.00
57.56
32.81
40.51
HyperGraphPro (Ours) (Ctxt.+Struct.)
59.38
66.04
39.06
49.91
Table 4: Performance comparison according to the retrieval methods on multi-hop question answering datasets using Qwen-2.5-3B. Ctxt. indicates the natural language context-based relevance. Struct. indicates the graph structure-based relevance. For this experiment, we do not apply progress-aware PO for HyperGraphPro for the fair comparison.
2Wiki.
HotpotQA
MuSiQue
NQ
Method
# Turns
F1
# Turns
F1
# Turns
F1
# Turns
F1
Qwen2.5-3B-Instruct
Graph-R1
2.63
57.56
3.65
56.75
3.79
40.51
1.21
44.75
HyperGraphPro ( ours )
2.51
69.76
3.14
67.59
2.70
52.26
1.17
46.84
Qwen2.5-7B-Instruct
Graph-R1
2.72
65.04
3.38
62.69
3.92
46.17
1.13
49.87
Table 5: Full comparison on the number of turns between Graph-R1 and HyperGraphPro (Ours).
Question & Answer
Graph-R1
HyperGraphPro (Ours)
Question: Who is the chief minister in the state where Anjarle is located? Answer: Devendra Fadnavis
⟨ think ⟩ To find out who the chief minister is in the state where Anjarle is located, we first need to identify which state Anjarle is in. ⟨ /think ⟩⟨ query ⟩ which state is Anjarle located in? ⟨ /query ⟩⟨ knowledge ⟩ Anjarle is a village in the Dapoli taluka of Ratnagiri district in the Maharashtra state of India. (rank 1) … ⟨ /knowledge ⟩⟨ think ⟩ From the provided knowledge base, Anjarle is located in the Ratnagiri district of Maharashtra state. Now we need to find out who the chief minister is in the state of Maharashtra. ⟨ /think ⟩⟨ query ⟩ who is the chief minister of Maharashtra state? ⟨ /query ⟩⟨ knowledge ⟩ Rank State Tax Revenues (INR Billions) 2010 - 2015 – India. (rank 1) … Prithviraj Chavan of the Congress party was the last Chief Minister of Maharashtra under the Congress/NCP alliance. (rank 7) … ⟨ /knowledge ⟩⟨ think ⟩ According to the information provided, the chief minister of Maharashtra state is Prithviraj Chavan of the Congress party. ⟨ /think ⟩⟨ answer ⟩ Prithviraj Chavan ⟨ /answer ⟩
⟨ think ⟩ To find the chief minister in the state where Anjarle is located, I first need to determine the state where Anjarle is situated. I can use a search query to find this information. ⟨ /think ⟩⟨ query ⟩ where is Anjarle located? ⟨ /query ⟩⟨ knowledge ⟩ Anjarle is a village in the Dapoli taluka of Ratnagiri district in the Maharashtra state of India. (rank 1) … ⟨ /knowledge ⟩⟨ think ⟩ From the knowledge base, I found that Anjarle is a village in the Dapoli taluka of Ratnagiri district in the Maharashtra state of India. Now, I need to find the chief minister of Maharashtra. ⟨ /think ⟩⟨ query ⟩ who is the chief minister of Maharashtra state? ⟨ /query ⟩⟨ knowledge ⟩ Prithviraj Chavan of the Congress party was the last Chief Minister of Maharashtra under the Congress/NCP alliance. (rank 1) … Devendra Fadnavis is the incumbent CM since 31 October 2014, the first from the BJP. (rank 8) With the exceptions of Manohar Joshi (Shiv Sena), Narayan Rane (Shiv Sena) and Devendra Fadnavis (Bharatiya Janata Party), all other CMs have been from the Congress or its breakaway parties. (rank 9) … ⟨ /knowledge ⟩⟨ think ⟩ From the knowledge base, I found that the chief minister of Maharashtra is Devendra Fadnavis. ⟨ /think ⟩⟨ answer ⟩ Devendra Fadnavis ⟨ /answer ⟩
Table 6: Comparison on outputs generated by Graph-R1 and HyperGraphPro (Ours) with Qwen2.5-3B.