Graph Retrieval-Augmented Generation (GraphRAG) has remarkably enhanced large language models on complex reasoning by leveraging structured entity topologies. However, existing frameworks heavily rely on standard autoregressive language models where the nature of inherent sequential generation severely hinders overall inference efficiency. Inspired by Diffusion Language Models (DLMs) that offer massive parallelism via continuous refine-in-parallel decoding, we aim to accelerate GraphRAG in the discrete space. However, it remains non-trivial for two challenges. First, partially denoised drafts are highly dynamic and uncertain, making dynamic graph grounding non-trivial. Second, raw denoising states are inherently noisy and unstable, making synchronous graph retrieval and multi-hop aggregation computationally prohibitive. To this end, we present LADDER, a novel framework that bridges diffusion language modeling with GraphRAG through graph-guided parallel decoding. Specifically, (i) we propose an event-driven self-clocking retrieval, inspired by our key insight that 88% of target entities emerge early in the partially denoised state, leading final commitment by an average of 5.7-9.6 steps. This mechanism dynamically triggers graph retrieval only when the set of graph-linkable entities expands, yielding an asynchronous self-clocking policy that bypasses learned gates or heuristic thresholds. (ii) An incomplete-query graph propagation module is designed to process the newly emerging entity queries using a specialized graph foundation model, continuously aggregating multi-hop evidence to sharpen parallel predictions and accelerate overall decoding convergence. Extensive experiments on three challenging multi-hop QA benchmarks show that LADDER raises average exact match from 39.6% to 45.2% while achieving a 4.1x latency reduction.
Figures & tables
Figure 1: Entities surface before their spans commit, early enough to be useful. (a) Fraction of entities still uncommitted x steps after surfacing in Et . (b) Fraction of final entities already detected after a given trajectory fraction. (c) Fraction of triggers with all gold documents already in memory; labels give the mean remaining steps. Section 4.4 targets the coverage loss visible here on long reasoning chains.
Figure 2: Overview of Ladder . The DLM exposes a partially denoised response state at each step. Newly matched graph entities trigger GFM propagation and retrieve top- K evidence, which conditions subsequent denoising updates. Equation 12 defines the complete retriever-training objective.
HotpotQA
2WikiMultihopQA
MuSiQue
Method
R@2
R@5
R@7
R@2
R@5
R@7
R@2
R@5
R@7
Autoregressive agent (Qwen2.5-7B)
RAPTOR ( Sarthi et al., 2024 )
30.1
56.8
65.2
43.8
66.1
73.5
17.6
31.9
39.1
LightRAG ( Guo et al., 2025 )
28.4
54.1
62.8
41.6
63.4
71.1
16.1
29.8
36.9
GraphRAG ( Edge et al., 2025 )
26.7
51.2
60.1
39.5
60.6
68.4
14.9
27.6
34.7
HippoRAG 2 ( Gutiérrez et al., 2025a )
32.6
59.6
68.1
46.7
70.8
77.9
21.6
37.6
41.8
Table 1: End-to-end retrieval recall (%). Autoregressive baselines and GFM-RAG (IRCoT) use multi-round thought-level retrieval; LADDER uses event-driven refreshes within one denoising trajectory. Bold marks the best value in each column.
HotpotQA
2WikiMultihopQA
MuSiQue
Average
Method
EM
F1
Time ↓
EM
F1
Time ↓
EM
F1
Time ↓
EM
F1
Time ↓
Autoregressive agent (Qwen2.5-7B)
RAPTOR
41.0
46.9
6.35
58.3
63.5
6.72
15.3
27.0
7.18
38.2
45.8
6.75
LightRAG
36.5
43.1
7.46
54.2
59.8
7.83
13.8
23.8
8.51
34.8
42.2
7.93
GraphRAG
33.7
40.6
9.72
51.4
56.9
10.16
12.5
21.9
11.24
32.5
39.8
10.37
HippoRAG 2
43.8
50.1
10.86
62.4
67.2
11.42
17.6
30.4
12.67
41.3
49.2
11.65
Table 2: End-to-end multi-hop QA performance. Autoregressive baselines and GFM-RAG (IRCoT) use multi-round thought-level retrieval; LADDER uses event-driven refreshes within one denoising trajectory. EM and F1 are percentages; time is per-question wall-clock seconds ( ↓ ). Bold marks the best value in each column.
Figure 3: Main results. Top: average exact match versus average per-question latency across the three datasets (Table 2 ), quality increasing upward and latency leftward. Middle: quality, latency, and convergence steps as a function of the commitment threshold τc (the vertical line marks τc=0.85 ). Bottom: convergence under four evidence conditions on 2WikiMultihopQA, with (left) positions committed per denoising step and (right) the distribution of steps to convergence.
Table 6
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
#Entities
Mean
Median
P90
Lead>0
2WikiMultihopQA
1,892
5.71
3.0
15
88.8%
HotpotQA
3,090
9.35
5.0
24
88.0%
MuSiQue
4,524
9.62
6.0
24
88.5%
Appendix
Table 5: Entity span-level lead time, in denoising steps, over 1,000 test questions per dataset. An entity is committed when all of its token positions are independently finalised by the sampler. Reproduced across nine shards (three per dataset), where the Lead>0 rate stays within 85.8% – 88.8% .
Dataset
Triggers/q
Useful-trigger rate
HotpotQA
3.17
53.3%
2WikiMultihopQA
2.07
79.0%
MuSiQue
4.35
42.6%
Appendix
Table 6: Trigger quality across benchmarks. Triggers/q is the mean number of real graph retrievals per question; Useful-trigger rate is the fraction, among triggers before full gold coverage, that increase the number of gold documents in the top- K list.
HotpotQA
2WikiMultihopQA
MuSiQue
Seed policy
R@7
EM
R@7
EM
R@7
EM
New-seed only ( ΔEt )
69.1
46.8
82.7
63.9
39.2
17.1
Cumulative ( Et )
72.9
49.4
86.9
67.4
41.4
18.8
Appendix
Table 7: Cumulative versus new-only seeding of the GNN start mask, per dataset, with the trigger and all other settings fixed. Cumulative ( Ladder ’s default) seeds every entity confirmed so far, Et ; new-seed uses only the entities that triggered the current event, ΔEt .
HotpotQA
2WikiMultihopQA
MuSiQue
Schedule
Time ↓
R@7
EM
F1
Time ↓
R@7
EM
F1
Time ↓
R@7
EM
F1
Retrieve-once ( t=0 )
2.28
61.9
42.1
51.8
1.75
71.2
57.6
62.4
3.62
33.6
13.9
24.1
Fixed, every 10 steps
2.41
66.8
45.4
55.1
1.85
76.8
61.2
66.3
3.80
36.2
15.4
26.3
Fixed, every 5 steps
2.67
69.7
47.2
57.0
2.05
80.5
63.8
69.1
4.21
38.5
16.8
28.2
Per-step
3.94
72.4
48.8
58.9
3.02
84.1
66.2
72.0
6.18
41.0
18.4
30.3
Random, firing rate 15%
2.52
66.1
44.8
54.5
1.94
76.1
60.6
65.7
3.97
36.8
15.9
26.9
Appendix
Table 8: Single-trajectory refresh schedules under the same Dream-7B backbone and full event-conditioned GFM retriever. Time is per-question end-to-end wall-clock seconds ( ↓ ). These ablations isolate refresh policy and use a different execution unit from the multi-round IRCoT baseline in Table 2 . Bold marks the best value in each column.
HotpotQA
2WikiMultihopQA
MuSiQue
Objective
R@7
R disc
EM
F1
R@7
R disc
EM
F1
R@7
R disc
EM
F1
(1)
Lretrieve only
66.4
42.7
45.2
54.0
78.9
50.2
63.4
68.5
40.6
26.1
17.2
30.0
(2)
+Llookahead
70.8
53.1
48.1
57.8
84.2
62.8
66.1
71.8
41.1
33.4
18.3
30.4
(3)
Full ( Ladder )
72.9
55.8
49.4
59.6
86.9
65.7
67.4
73.5
41.4
34.9
18.8
30.7
(4)
Full −Lanchor
71.2
54.2
48.3
58.1
85.1
64.0
66.3
72.1
39.8
32.7
17.9
29.6
Appendix
Table 9: Training objectives, per dataset. Bold marks the best value in each column.
Dataset
K
EM
F1
R@ K (final)
Triggers/q
Time ↓
2WikiMultihopQA
5
59.0
64.6
80.1
2.07
1.56
7
67.4
73.5
86.9
2.07
1.88
10
73.0
76.3
92.1
2.08
2.70
15
71.0
75.7
93.9
2.09
3.55
MuSiQue
5
15.0
25.1
37.2
4.14
3.17
7
18.8
30.7
41.4
4.35
3.95
Appendix
Table 10: Sensitivity of Ladder to the retrieval budget K on 2WikiMultihopQA and MuSiQue, with the generator, retriever, and trigger policy held fixed. EM and F1 are percentages; R@ K is the final-state recall at the evaluated budget; Triggers/q is the mean number of real graph retrievals per question; latency is per-question wall-clock seconds ( ↓ ).
Dataset
Evidence
Mean
SD
Median
Paired
2WikiMultihopQA
No retrieval
41.70
15.68
36.0
—
Random passages
24.05
11.30
23.0
—
Graph retrieval
15.24
12.57
11.0
—
Oracle evidence
12.76
8.43
10.0
97%
HotpotQA
No retrieval
37.91
11.37
37.0
—
Random passages
31.00
13.54
31.0
—
Appendix
Table 11: Steps to convergence by evidence condition. Paired is the share of questions on which oracle evidence converges in strictly fewer steps than no retrieval.
Figure 4: Convergence on HotpotQA, in the format of Figure 3 .
Step
Event
Retrieval
Denoising state ut
Gold
0
—
fwd
(question only)
1/2
1
{college, Oxford, University}
fwd
Charlotte Brewer is at H Soulsford College, Oxford. (2): …
1/2
2
—
✓
Charlotte Brewer is at professor offord College, Oxford. (2): …
1/2
3
—
✓
(wording drifts further; entity set unchanged)
1/2
4 – 5
—
∅
(proxy text identical to step 3 )
1/2
6
—
✓
Charlotte Brewer is at professor offord College, Oxford. (2): … (still wrong)
1/2
Appendix
Table 12: Case study: “What larger educational institution is the university where Charlotte Brewer went, a part of?” Gold documents: Charlotte Brewer and Hertford College, Oxford . Bold tokens in the denoising state are independently committed positions; the remainder is the proxy’s argmax fill of still-masked positions. GNN marks a forward pass; ✓ marks a scheduled retrieval call served by cached graph scores; and ∅ marks a step where the unchanged query skips the retrieval call. Gold is the number of gold documents in the retrieved top- K out of the total.
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.
Retrieval-Augmented Generation (RAG) mitigates hallucinations in Large Language Models (LLMs) by grounding the generation process on external knowledge. However, standard RAG approaches struggle with multi-hop reasoning. While recent graph-based RAG methods improve the retrieval of interconnected chunks, they often rely on computationally expensive and error-prone LLM-based extraction pipelines. To address these issues, we propose TIGRAG (Token-Induced GraphRAG), an efficient graph-augmented RAG framework based on a token co-occurrence Knowledge Graph. TIGRAG directly models topological relationships between tokens using sliding-window co-occurrence statistics, thus enabling scalable graph construction. During inference, it combines graph-based semantic expansion and neural reranking to retrieve interconnected evidence for multi-hop reasoning. Specifically, it introduces an iterative entity-driven retrieval strategy that progressively expands the query using bridging entities extracted from previously retrieved contexts. We evaluated TIGRAG on three widely adopted multi-hop Question Answering (QA) benchmarks. Experimental results demonstrated that our framework consistently outperforms dense retrieval and graph-based RAG methods in both retrieval and downstream QA tasks, while substantially reducing indexing time, inference latency, and prompt footprint.
Gianluca Bonifazi, Christopher Buratti, Michele Marchetti +5
aUniversità Politecnica delle Marche, Ancona, Italy · bUniversità di Modena e Reggio Emilia, Modena, Italy
Retrieval-Augmented Generation (RAG) has become a core paradigm for enhancing factual grounding and multi-hop reasoning in Large Language Models (LLMs). Traditional text-based RAG often retrieves logically irrelevant pseudo-evidence, while graph-based RAG is frequently hindered by search-time pruning, which may discard potentially valid reasoning paths. Existing hybrid approaches primarily adopt simple evidence concatenation or unidirectional enhancement, which fails to address the fundamental "Information Island" problem caused by asymmetric reasoning flows between unstructured text and structured graphs. We propose TGS-RAG, a unified framework for Text-Graph Synergistic enhancement. TGS-RAG introduces a bidirectional mechanism: (i) a Graph-to-Text channel that employs a Global Voting strategy from visited graph nodes to re-rank and refine textual evidence, filtering out semantic noise; and (ii) a Text-to-Graph channel that utilizes the Memory-based Orphan Entity Bridging algorithm. This algorithm utilizes textual cues to proactively resurrect valid but previously pruned reasoning paths from the search history without additional database overhead. Experimental results on multiple multi-hop reasoning benchmarks demonstrate that TGS-RAG significantly outperforms state-of-the-art baselines, achieving a superior balance between retrieval precision and computational efficiency.
Jiarui Zhong, Hong Cai Chen
School of Automation, Southeast University, Nanjing 210096, China