Organizations: School of Computer Science and Technology, Soochow University, China · Soochow University, China · Hong Kong University of Science and Technology, China · Computer Science and Engineering, The Chinese University of Hong Kong, China · DSA, The Hong Kong University of Science and Technology, China · School of Education, Zhejiang Normal University, China
The trustworthiness of Web content faces a growing threat from multimodal fake news. Real news usually shows aligned image-text semantics, while fake news often contains a few conflicts at varying levels, such as mismatches in objects, scenes, events, and claims. However, existing methods usually treat consistency and conflict in a coarse and static way, limiting their ability to capture sparse conflicts and adapt to varying conflict levels. Inspired by the distinct knowledge encoded across layers in large language models, we propose Multi-Level Consistency and Conflict Decoupling (MLCCD), which captures conflicts at different levels from different layers of multimodal large language models (MLLMs). Specifically, we first use the MLLM's final decision representation to select visual and textual tokens relevant to the task, preventing irrelevant content from diluting sparse conflicts. Then, we preserve tokens whose representations change sharply between adjacent layers, capturing the knowledge each layer adds. Finally, we decouple this knowledge into consistency and conflict evidence and aggregate each across layers with its own router, adapting to varying conflict levels. Experiments on three benchmarks show that MLCCD outperforms existing methods.
Figures & tables
Figure 1. Illustration of two challenges in multimodal fake news detection. (a) Sparse conflicts can be diluted when aggregated with other content. (b) The importance of consistency and conflict evidence can vary across samples and semantic levels.
Figure 2. Overview of MLCCD. The frozen LoRA tuned MLLM provides layer wise hidden states, and the proposed evidence network selects task relevant tokens, extracts discriminative layer knowledge, and adaptively routes consistency and conflict evidence for fake news prediction.
Dataset
Train
Dev
Test
Lang.
Weibo
7461
–
1929
Chinese
Pheme
1412
221
385
English
AMG
3268
501
883
English
Table 1. Dataset statistics.
Variant
Acc.
F1-Ma
F1-R
F1-F
w/o TRTS
92.21
90.45
94.55
86.36
w/o DLKE
92.21
90.70
94.44
86.96
w/o ACCR
92.47
90.70
94.76
86.64
Shared Router
91.69
89.46
94.31
84.62
w/o Dis.
92.73
91.09
94.91
87.27
MLCCD
93.25
91.73
95.27
88.18
Table 2. Ablation results on Pheme. All values are percentages. Best results are in bold. F1-R and F1-F denote real class F1 and fake class F1.
Category
Method
AMG
PHEME
Weibo
Acc.
F1-Ma
F1-R
F1-F
Acc.
F1-Ma
F1-R
F1-F
Acc.
F1-Ma
F1-R
F1-F
General fake news detection methods
CFFN ( Li et al., 2023a )
82.90
82.39
85.38
79.40
82.86
79.22
87.91
70.54
89.22
89.21
89.52
88.90
MTS ( Sun et al., 2025 )
85.05
84.51
87.40
81.62
89.61
87.72
92.54
82.91
89.58
89.54
90.15
88.94
MIMoE-FND ( Liu et al., 2025 )
84.03
83.64
86.16
81.12
88.57
86.62
91.73
81.51
89.89
89.88
90.19
89.58
DDU ( Chen et al., 2026 )
85.39
84.91
87.61
82.21
86.49
83.88
90.37
77.39
90.93
90.93
91.02
90.83
General multimodal methods
DMD ( Li et al., 2023b )
84.60
84.25
86.59
81.91
88.57
85.24
92.25
78.22
89.42
89.41
89.75
89.08
Table 3. Overall results on AMG, PHEME, and Weibo with seed 42. All values are percentages. Acc. and F1-Ma denote accuracy and macro F1, respectively. F1-R and F1-F denote real class F1 and fake class F1, respectively.
Figure 3. Hyperparameter sensitivity of MLCCD. (a) Text keep ratio ρqt on Weibo, AMG, and Pheme, shown on a logarithmic horizontal axis. (b) Image keep ratio ρqv , (c) token keep ratio ρe in DLKE, and (d) number of routed layers on Pheme. Outer circles indicate the default settings. Panels (b)–(d) share the same vertical scale. Each point represents one reported experimental result, and lines connect the evaluated settings.
Sample type
N
Qwen3-VL-8B + LoRA
MLCCD
Δ (pp)
Correct
Rate (%)
Correct
Rate (%)
Counterfactual samples: detection as fake
Object conflict
100
23
23.00
50
50.00
+27.00
Scene conflict
100
23
23.00
52
52.00
+29.00
Event conflict
100
14
14.00
34
34.00
+20.00
Claim conflict
100
14
14.00
33
33.00
+19.00
Table 4. Results on the controlled counterfactual evaluation set. N denotes the number of samples, and Correct denotes the number assigned the intended label. Rates are percentages: counterfactual detection rates in the upper block and correct acceptance rates for real samples in the lower block. Δ denotes the rate of MLCCD minus that of the LoRA baseline in percentage points.
Edit ratio
N
LoRA
MLCCD
redit≤5%
151
23 (15.23)
54 (35.76)
5%<redit≤10%
153
29 (18.95)
65 (42.48)
10%<redit≤20%
96
22 (22.92)
50 (52.08)
All samples
400
74 (18.50)
169 (42.25)
Table 5. Counterfactual detection results grouped by edit ratio. N denotes the number of samples. Each method entry reports the number of detected counterfactuals, followed by the detection rate (%) in parentheses. LoRA denotes Qwen3-VL-8B + LoRA.
Image replacement
N
Original
Replaced
Uniform gray
400
4.84
0.96
Donor image
256
4.70
1.02
Table 6. Effect of image replacement on MLCCD. N denotes the number of counterfactual and control pairs. Original and Replaced report the mean paired score difference under the respective image conditions, evaluated on the same pairs in each row.
Method
Total training time (s)
Training overhead
Inference time (s)
Qwen3-VL-8B + LoRA
845
–
≈45
MLCCD
1102
+257 s (+30.4%)
≈48
Table 7. Runtime comparison on Pheme. MLCCD training time includes both LoRA adaptation and evidence network training. Inference time covers all 385 test samples. Timings are derived from execution logs and include data handling and logging overhead.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
ρqt
Weibo
AMG
Pheme
Acc.
F1-Ma
F1-R
F1-F
Acc.
F1-Ma
F1-R
F1-F
Acc.
F1-Ma
F1-R
F1-F
0.0125
95.28
95.28
95.39
95.17
88.45
88.26
89.74
86.79
91.43
89.29
94.08
84.51
0.025
95.28
95.28
95.36
95.20
90.37
90.07
91.80
88.34
91.69
89.76
94.20
85.32
0.05
95.33
95.33
95.45
95.22
89.47
89.17
90.96
87.38
92.21
90.24
94.62
85.85
0.1
95.23
95.23
95.29
95.17
88.79
88.57
90.15
86.99
93.25
91.73
95.27
88.18
0.3
95.28
95.28
95.34
95.22
88.67
88.45
90.06
86.84
92.73
91.04
94.93
87.16
Appendix
Table 8. Sensitivity to the text keep ratio ρqt . All performance scores are percentages. The default ratios are 0.3 , 0.025 , and 0.1 for Weibo, AMG, and Pheme, respectively. F1-R and F1-F denote real class F1 and fake class F1.
Parameter
Value
Acc.
F1-Ma
F1-R
F1-F
ρqv
0.1
92.73
91.04
94.93
87.16
0.3†
93.25
91.73
95.27
88.18
0.5
91.69
89.52
94.29
84.76
ρe
0.3
91.69
89.46
94.31
84.62
0.5†
93.25
91.73
95.27
88.18
0.7
92.47
90.75
94.74
86.76
Appendix
Table 9. Sensitivity to the image keep ratio ρqv , the token keep ratio ρe in DLKE, and the number of routed layers on Pheme. All performance scores are percentages. † marks the default setting in each group. F1-R and F1-F denote real class F1 and fake class F1.
Method
Acc.
Fake News
Real News
Prec.
Rec.
F1
Prec.
Rec.
F1
CFFN
82.90
74.42
85.09
79.40
89.63
81.52
85.38
MTS
85.05
77.93
85.67
81.62
90.34
84.66
87.40
MIMoE-FND
84.03
74.81
88.60
81.12
91.84
81.15
86.16
DDU
85.39
77.81
87.13
82.21
91.20
84.29
87.61
DMD
84.60
75.12
90.06
81.91
92.81
81.15
86.59
Appendix
Table 10. Detailed classification results on AMG.
Method
Acc.
Fake News
Real News
Prec.
Rec.
F1
Prec.
Rec.
F1
CFFN
82.86
71.17
69.91
70.54
87.59
88.24
87.91
MTS
89.61
80.17
85.84
82.91
93.94
91.18
92.54
MIMoE-FND
88.57
77.60
85.84
81.51
93.85
89.71
91.73
DDU
86.49
76.07
78.76
77.39
91.04
89.71
90.37
DMD
88.57
88.76
69.91
78.22
88.51
96.32
92.25
Appendix
Table 11. Detailed classification results on Pheme.
Method
Acc.
Fake News
Real News
Prec.
Rec.
F1
Prec.
Rec.
F1
CFFN
89.22
88.52
89.28
88.90
89.88
89.16
89.52
MTS
89.58
91.40
86.60
88.94
88.04
92.37
90.15
MIMoE-FND
89.89
89.34
89.82
89.58
90.41
89.96
90.19
DDU
90.93
88.83
92.93
90.83
93.07
89.06
91.02
DMD
89.42
88.98
89.17
89.08
89.84
89.66
89.75
Appendix
Table 12. Detailed classification results on Weibo.
Multimodal fake news detection is crucial for mitigating societal disinformation. Existing approaches attempt to address this by fusing multimodal features or leveraging Large Language Models (LLMs) for advanced reasoning. However, these methods suffer from serious limitations, including a lack of comprehensive multi-view judgment and fusion, and prohibitive reasoning inefficiency due to the high computational costs of LLMs. To address these issues, we propose LLM-Guided Multi-View Reasoning Distillation for Fake News Detection ( LLM-MRD), a novel teacher-student framework. The Student Multi-view Reasoning module first constructs a comprehensive foundation from textual, visual, and cross-modal perspectives. Then, the Teacher Multi-view Reasoning module generates deep reasoning chains as rich supervision signals. Our core Calibration Distillation mechanism efficiently distills this complex reasoning-derived knowledge into the efficient student model. Experiments show LLM-MRD significantly outperforms state-of-the-art baselines. Notably, it demonstrates a comprehensive average improvement of 5.19% in ACC and 6.33% in F1-Fake when evaluated across all competing methods and datasets. Our code is available at https://github.com/Nasuro55/LLM-MRD
Weilin Zhou, Shanwen Tan, Enhao Gu +1
Xinjiang University, Urumqi, China · Joint International Research Laboratory of Silk Road Multilingual Cognitive Computing, Xinjiang University, Urumqi, China · Sichuan University, Chengdu, China
In recent years, multimodal multidomain fake news detection has garnered increasing attention. Nevertheless, this direction presents two significant challenges: (1) Failure to Capture Cross-Instance Narrative Consistency: existing models usually evaluate each news in isolation, fail to capture cross-instance narrative consistency, and thus struggle to address the spread of cluster based fake news driven by social media; (2) Lack of Domain Specific Knowledge for Reasoning: conventional models, which rely solely on knowledge encoded in their parameters during training, struggle to generalize to new or data-scarce domains (e.g., emerging events or niche topics). To tackle these challenges, we introduce Retrieval-Augmented Multimodal Model for Fake News Detection (RAMM). First, RAMM employs a Multimodal Large Language Model (MLLM) as its backbone to capture cross-modal semantic information from news samples. Second, RAMM incorporates an Abstract Narrative Alignment Module. This component adaptively extracts abstract narrative consistency from diverse instances across distinct domains, aggregates relevant knowledge, and thereby enables the modeling of high-level narrative information. Finally, RAMM introduces a Semantic Representation Alignment Module, which aligns the model's decision-making paradigm with that of humans - specifically, it shifts the model's reasoning process from direct inference on multimodal features to an instance-based analogical reasoning process. Extensive experimental results on three public datasets validate the efficacy of our proposed approach. Our code is available at the following link: https://github.com/li-yiheng/RAMM
Yiheng Li, Weihai Lu, Hanyi Yu +1
University of International Business and Economics · Beijing, China · Peking University +4
The rapid rise of generative AI has made multimodal fake news increasingly realistic and pervasive, posing severe threats to public trust and social stability. Existing detection methods rely heavily on manipulation-specific models and large-scale labeled data, resulting in poor generalization to emerging manipulation types. We observed that the essence of manipulated misinformation lies in its intrinsic conflicts, i.e., semantic or physical inconsistencies either across modalities or with common world knowledge. Inspired by this observation, we propose Conflict-Oriented REasoning (CORE) framework, an effective paradigm that learns to endows multimodal large language models (MLLMs) with explicit conflict-capturing capability. To this end, CORE first constructs the Conflict Attribution Corpus (CAC) with fine-grained annotations of conflict factors and sources, providing essential data support for subsequent conflict perception training. By performing conflict-oriented representation enhancement and reasoning based on CAC, CORE achieves robust and generalizable conflict detection, effectively and rapidly adapting to unseen manipulation types with a few samples or in even zero-shot settings. Extensive experiments demonstrate that CORE surpasses state-of-the-art models. The dataset and code are publicly available at https://github.com/shen8424/CORE.
Jinjie Shen, Yaxiong Wang, Yujiao Wu +5
School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China · Wuhan University, Wuhan, China · Lab for Intelligence and visiON (LION) +2