Multimodal graph learning requires jointly training over graph structure and heterogeneous node attributes, yet existing methods largely decouple these processes: prior multimodal graph neural networks (GNNs) focus on aligning modalities in a shared embedding space while operating on fixed or weakly adapted graph structures, and graph structure learning approaches infer topology from unimodal node representations without accounting for multimodal interactions. This separation fundamentally limits the ability of GNNs to capture semantically meaningful relationships in multimodal settings, where observed edges are often noisy, incomplete, or misaligned with underlying semantics. We propose ReCoG (Reciprocal Co-Evolution for Multimodal Graph Learning), a new learning paradigm that tightly couples graph structure learning and multimodal representation learning through end-to-end reciprocal interaction. Concretely, ReCoG integrates (i) a multimodal graph refiner that infers and corrects edges using cross-modal semantic evidence, and (ii) a coupled cross-modal message passing mechanism that performs joint intra- and inter-modality propagation over the refined graph. This unified design yields greater expressiveness than decoupled or two-stage formulations and allows dynamic interaction between topology and representation learning. Across diverse benchmarks for node classification and link prediction, ReCoG consistently outperforms strong multimodal graph structure learning baselines, including graph foundation models. Our results demonstrate that reciprocal co-evolution of structure and semantics is important for effective multimodal graph learning, challenging the prevailing separation between topology and representation learning.
Figures & tables
Figure 1: Overview of the proposed ReCoG framework. Left: The input multimodal graph with text and vision features. Center-top: The Structure Refiner module with two variants: a similarity-based version, and a residual MLP-based version. Center-bottom: Cross-modal coupled message passing over L layers, where each layer performs intra-modal neighbor aggregation and cross-modal injection. Right: Modality fusion followed by a task-specific head. The entire pipeline is trained end-to-end. Best viewed in magnification. See text for detail.
Figure 2: Topology–semantics alignment before and after refinement. Left : distribution of multimodal cosine similarity for connected and disconnected node pairs in the original graph (RedditM). Right : similarity distribution of edges added by the cosine refiner compared to disconnected pairs.
Ele-fashion
Goodreads-NC
Model
CLIP
T5+ViT
IBind
T5+DINO
CLIP
T5+ViT
IBind
T5+DINO
MMGCN
86.10 ± 0.50
82.39 ± 0.30
86.21 ± 0.94
85.53 ± 0.33
83.29 ± 0.20
81.85 ± 0.22
80.58 ± 1.08
82.44 ± 0.11
MGAT
84.66 ± 0.29
84.01 ± 0.08
86.12 ± 0.08
84.54 ± 0.27
76.48 ± 0.59
75.43 ± 0.76
69.45 ± 6.25
74.98 ± 1.23
GCN
79.83 ± 0.03
79.63 ± 0.07
80.35 ± 0.02
79.37 ± 0.04
81.61 ± 0.01
81.67 ± 0.03
78.91 ± 0.04
81.71 ± 0.03
SAGE
87.10 ± 0.02
84.41 ± 0.09
87.71 ± 0.13
85.31 ± 0.09
83.30 ± 0.02
83.03 ± 0.04
80.39 ± 0.21
82.99 ± 0.08
MLP
85.16 ± 0.03
84.98 ± 0.05
88.73 ± 0.01
84.87 ± 0.01
72.29 ± 0.02
67.82 ± 0.07
58.75 ± 0.05
68.83 ± 0.03
Table 1: Node classification accuracy (%) on the MM-Graph [ 38 ] . Best and runner-up are highlighted.
Movies
Toys
Grocery
Reddit-S
Reddit-M
Model
L+C
QWen
LLaVL
L+C
QWen
LLaVL
L+C
QWen
LLaVL
L+C
QWen
LLaVL
L+C
QWen
LLaVL
GCN
53.98
55.19
54.92
81.59
80.95
81.57
84.64
82.96
84.17
89.77
90.88
91.13
71.42
72.96
75.64
SAGE
56.08
58.01
57.39
82.35
81.45
82.35
86.65
85.55
86.81
91.08
92.57
92.94
80.84
81.23
85.61
sim
ReCoG-GCN
56.97
57.72
57.51
81.98
81.20
82.19
85.89
83.46
85.03
97.20
96.98
97.26
90.55
90.24
90.87
ReCoG-SAGE
56.94
58.92
60.00
82.58
82.27
82.75
86.86
86.24
89.05
96.76
96.38
96.73
88.75
89.18
90.11
res
ReCoG-GCN
57.45
60.12
58.62
82.07
82.27
82.27
86.18
86.39
86.56
96.89
96.82
96.73
87.28
87.31
87.42
Table 2: Node classification accuracy (%) on the MAGB datasets [ 34 ] .
Table 4: Ablation study on node classification accuracy (%). † MMGCN uses separate GCNs for each modality with late fusion.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Encoder
Amazon-Sports
Amazon-Cloth
MRR
H@1
H@10
MRR
H@1
H@10
MMGCN
CLIP
31.96 ± 0.10
16.35 ± 0.11
68.46 ± 0.08
22.20 ± 0.05
10.76 ± 0.10
46.62 ± 0.12
ViT+T5
30.33 ± 0.08
15.01 ± 0.05
66.41 ± 0.11
19.45 ± 0.34
9.22 ± 0.20
40.49 ± 0.61
ImageBind
31.74 ± 0.21
16.45 ± 0.13
67.39 ± 0.74
24.72 ± 0.19
12.47 ± 0.09
51.32 ± 0.56
DINO+T5
30.04 ± 0.27
14.98 ± 0.07
64.56 ± 0.56
21.77 ± 0.23
10.47 ± 0.12
45.81 ± 0.52
MGAT
CLIP
27.56 ± 0.30
13.55 ± 0.29
60.21 ± 0.21
21.38 ± 0.23
10.39 ± 0.22
44.60 ± 0.36
Appendix
Table 5: Link prediction performance on Amazon-Sports and Amazon-Cloth. [ 38 ] . The Best and runner-up are highlighted.
Refiner
K
Node Classification (%)
Link Prediction (MRR %)
Ele-Fashion
Goodreads
Amazon-Sports
Amazon-Clothing
sim
1
87.60 ± 0.05
89.65 ± 0.05
36.72 ± 0.11
27.07 ± 0.10
3
88.17 ± 0.03
90.43 ± 0.09
37.81 ± 0.06
27.93 ± 0.07
5
88.56 ± 0.05
90.31 ± 0.05
38.53 ± 0.03
28.30 ± 0.03
8
88.53 ± 0.05
90.16 ± 0.06
38.32 ± 0.09
28.01 ± 0.06
12
87.95 ± 0.06
89.43 ± 0.10
37.23 ± 0.10
26.68 ± 0.05
Appendix
Table 6: Ablation study on the number of candidate neighbors K in the graph refiner.
Figure 3: Topology–semantics alignment.
Component
Complexity
Cosine similarity ( ST , SV )
O(BNd)
Original edge scoring (MLP)
O(E⋅d⋅dh)
Coarse filtering
O(BNdh)
Fine scoring (MLP)
O(Brk⋅d⋅dh)
Cross-modal message passing ( ×L )
O(L(E⋅dh+N⋅dh2))
Appendix
Table 7: Per-batch computational complexity of ReCoG major components.
Dataset
#Nodes
#Edges
#Classes
Task
Ele-fashion ‡
97,766
199,602
12
NC
Goodreads-NC ‡
685,294
7,235,084
10
NC
Movies †
16,672
218,390
20
NC
Toys †
20,695
126,886
18
NC
Grocery †
17,074
171,340
20
NC
Reddit-S †
15,894
566,160
20
NC
Appendix
Table 8: Dataset statistics. Datasets marked with † are from the MAGB benchmark [ 34 ] ; those marked with ‡ are from MM-Graph [ 38 ] . NC = node classification, LP = link prediction.
Multimodal-attributed graphs (MAGs) are a fundamental data structure for multimodal graph learning (MGL), enabling both graph-centric and modality-centric tasks. However, our empirical analysis reveals inherent topology quality limitations in real-world MAGs, including noisy interactions, missing connections, and task-agnostic relational structures. A single graph derived from generic relationships is therefore unlikely to be universally optimal for diverse downstream tasks. To address this challenge, we propose Task-aware Modality and Topology co-Evolution (TMTE), a novel MGL framework that jointly and iteratively optimizes graph topology and multimodal representations toward the target task. TMTE is motivated by the bidirectional coupling between modality and topology: multimodal attributes induce relational structures, while graph topology shapes modality representations. Concretely, TMTE casts topology evolution as multi-perspective metric learning over modality embeddings with an anchor-based approximation, and formulates modality evolution as smoothness-regularized fusion with cross-modal alignment, yielding a closed-loop task-aware co-evolution process. Extensive experiments on 9 MAG datasets and 1 non-graph multimodal dataset across 6 graph-centric and modality-centric tasks show that TMTE consistently achieves state-of-the-art performance. Our code is available at https://anonymous.4open.science/r/TMTE-1873.
Yinlin Zhu, Xunkai Li, Di Wu +3
Sun Yat-sen University, Guangzhou, China · Beijing Institute of Technology, Beijing, China
Multimodal graphs, where nodes contain heterogeneous features such as images and text, are increasingly common in real-world applications. Effectively learning on such graphs requires both adaptive intra-modal message passing and efficient inter-modal aggregation. However, most existing approaches to multimodal graph learning are typically extended from conventional graph neural networks and rely on static structures or dense attention, which limit flexibility and expressive node embedding learning. In this paper, we propose a novel multimodal graph representation learning framework with Dynamic information Pathways (DiP). By introducing modality-specific pseudo nodes, DiP enables dynamic message routing within each modality via proximity-guided pseudo-node interactions and captures inter-modality dependence through efficient information pathways in a shared state space. This design achieves adaptive, expressive, and sparse message propagation across modalities with linear complexity. We conduct the link prediction and node classification tasks to evaluate performance and carry out full experimental analyses. Extensive experiments across multiple benchmarks demonstrate that DiP consistently outperforms baselines.
Xiaobin Hong, Mingkai Lin, Xiaoli Wang +2
State Key Laboratory for Novel Software Technology, Nanjing University · Nanjing Forest University, College of Information Science and Technology · Zhejiang Provincial Seaport Investment & Operation Group Co. Ltd
Multimodal Graph Neural Networks (MGNNs) have shown strong potential for learning from multimodal attributed graphs, yet most existing approaches rely on tightly coupled architectures that suffer from prohibitive computational overhead. In this paper, we present a systematic empirical analysis showing that decoupled MGNNs are substantially more efficient and scalable for large-scale graph learning. However, we identify a critical bottleneck in existing decoupled pipelines, namely modal conflict, which arises in both the propagation and aggregation stages. Specifically, independent multi-hop diffusion causes cross-modal semantic divergence during propagation, while naive fusion fails to align multi-hop feature trajectories during aggregation, jointly limiting effective representation learning. To address this challenge, we propose CAMPA, a Cross-modal Aligned Multimodal Propagation & Aggregation framework for decoupled multimodal graph learning. Concretely, CAMPA introduces a two-stage alignment mechanism: (1) cross-modal aligned propagation, which injects cross-modal similarity priors into message passing to preserve semantic consistency without additional parameter overhead; (2) trajectory aligned aggregation, which leverages trajectory-level self-attention and cross-attention to capture and align long-range dependencies across modalities and hops. Extensive experiments on diverse benchmark datasets and tasks demonstrate that CAMPA consistently outperforms strong coupled and decoupled baselines while preserving the efficiency advantages of the decoupled paradigm.