MM-FusionNet: Context-Aware Dynamic Fusion for Multi-modal Fake News Detection with Large Vision-Language Models

Benjamin Turner; Jingyuan Zhao; Junhao He; Tianyu Liu

REVIEW 3 major objections 5 minor 27 references

MM-FusionNet: Context-Aware Dynamic Fusion for Multi-modal Fake News Detection with Large Vision-Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A context-aware gating module lets text and image cues each carry the decision when they conflict, pushing multi-modal fake-news detection to an F1 of 0.938 on an 80,000-sample dataset.

desk verdict The central SOTA claim doesn't stand up: the 'existing baselines' are the authors' own ablations, and the dataset is private. read the letter →

arxiv 2508.05684 v1 pith:TM4PFNPY submitted 2025-08-05 cs.CR cs.LG

Junhao He , Tianyu Liu , Jingyuan Zhao , Benjamin Turner This is my paper

classification cs.CRcs.LG

keywords multi-modalfakenewsdetectiondynamicmodalgatingcross-modalattentionlargevision-languagemodelscontext-awarefusionLMFNDdatasetmodelinterpretabilitymodalityrobustness

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a lightweight, learnable gating mechanism—one that assigns separate importance weights to text and image features after they have attended to each other—gives a small but real accuracy advantage over fixed-weight fusion for multi-modal fake news detection. The proposed MM-FusionNet uses a fine-tuned LLM for text, a CLIP-based ViT for images, and a Context-Aware Dynamic Fusion Module that scales each modality's representation with sigmoid weights before concatenating and classifying. On the paper's large-scale LMFND benchmark (80,000 samples), it reports an F1 score of 0.938, roughly 0.5% above the best fixed-weight cross-modal attention baseline and above both single-modality baselines. The authors also show the learned weights shift between text and image across samples, and that the model degrades gracefully when one modality is missing or noisy. A sympathetic reader would care because fake news detection in the wild frequently hinges on spotting which modality is lying, and an adaptive weighting rule is a direct way to encode that.

What carries the argument

Context-Aware Dynamic Fusion Module (CADFM): bi-directional cross-modal attention (text attends to image, image attends to text, each with residual connections) followed by a lightweight MLP gate with sigmoid outputs $\alpha_T$ and $\alpha_I$ that scale the attended features before final fusion and MLP classification. The gating network is the load-bearing piece: the paper's ablation shows removing it (leaving fixed-weight cross-modal attention) costs 0.5% F1, and the perturbation experiments show it also serves as a noise filter when inputs are missing or corrupted.

What would settle it

Re-run MM-FusionNet and the fixed-weight attention baseline on a publicly released version of LMFND (or another public multi-modal fake news dataset) with cross-validation, and check whether the gating module's F1 advantage over fixed-weight attention is consistently above chance-level variation. If the advantage vanishes or the dataset cannot be independently reproduced, the central claim is falsified.

Watch

Extended reading notes

Core claim

MM-FusionNet claims that the bottleneck in multi-modal fake news detection is not feature extraction but fusion: when text and image carry contradictory or imbalanced evidence, a fixed fusion rule leaves accuracy on the table. Its Context-Aware Dynamic Fusion Module first lets text and image features query each other in both directions, then a small MLP reads the concatenated attended features and outputs two sigmoid scalars $\alpha_T$ and $\alpha_I$. These scalars gate each modality before concatenation and classification. The paper reports that this dynamic gate raises F1 from 0.933 (its own fixed-weight attention baseline) to 0.938 on LMFND, while single-modal baselines sit at 0.917 (text

Load-bearing premise

The central claim rests on the assumption that the in-house LMFND dataset is correctly labeled and representative, and that the fixed-weight cross-modal attention baseline is the strongest plausible comparison; if either is not true, the measured 0.5% margin does not support the conclusion.

Editorial extensions

If this is right

If the reported margin holds up, any multi-modal detector that already uses cross-modal attention can gain roughly 0.5 F1 by adding a small gating MLP, at negligible parameter cost.
The gating weights give an interpretability hook: a deployed detector can summarize a decision as 'the model trusted the text' or 'the model trusted the image,' which is useful for social media moderation workflows.
The robustness results imply the same architecture can be served in degraded modes—missing image or stripped text—without retraining, since the gate learns to down-weight the absent modality.
Learned weights are context-dependent rather than fixed: 38.5% of test samples favor text, 25.1% favor image, and 36.4% are balanced, so the model is not just rediscovering a static weighting.
The near-human F1 on the annotated subset suggests model-assisted human review, rather than fully automated detection, is a realistic deployment pattern.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension the paper does not run: apply the same CADFM to a public multi-modal fake news benchmark with externally reported state-of-the-art results. That would show whether the 0.5% margin generalizes beyond the paper's own baselines.
The paper leaves open how the gate behaves under adversarial modality poisoning, e.g. a fake image that is also salient. A natural stress test is to train with one modality deliberately misleading and see whether the gating weights learn to ignore it; the current perturbation study only covers random noise and missing inputs.
The reported closeness to human-level performance on a 500-sample subset suggests a cheaper path to large-scale annotation: use low-confidence predictions or gating-weight disagreement to flag samples for human review.
Because the gate's advantage plausibly comes from handling text-image incongruity, integrating a dedicated inconsistency score rather than raw attended features might push accuracy further; the paper's error analysis hints at this but does not test it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the paper is a coherent application of standard cross-modal attention plus a gating MLP to fake news detection, with a clean internal ablation. But the abstract's claim that it 'surpasses existing multi-modal baselines by 0.5%' is not supported. The two multi-modal baselines in Table I — Simple Concatenation and Fixed-Weight Cross-Modal Attention — are built in this paper. Section IV-C-1 explicitly says the fixed-weight baseline is MM-FusionNet without the gating network. So the headline gain is against an internal ablation, not against SpotFake, EANN, MVAE, or the entity-oriented alignment network, none of which appear in the table. That's the load-bearing flaw.

What's good: the framework is clearly described, the equations check out, and the ablations isolate the contributions. The perturbation analysis (missing text/image, noise) is a nice addition, and the gating weight statistics show the mechanism does vary by input. The human evaluation is small but gives context. The writing is readable.

The soft spots beyond the central one: the LMFND dataset is described as newly collected with no public link, so nobody can reproduce or verify the 80k samples. There are no error bars or significance tests on the 0.5% F1 gap — with a private dataset, that gap could easily be noise. The novelty is modest: bi-directional cross-modal attention and sigmoid gating are both standard in multimodal fusion; the contribution is the combination applied to this task. The paper cites a lot of relevant work but doesn't benchmark against any of it, which is the more serious problem.

Who this is for: readers interested in fusion architectures for misinformation could glance at the method, but anyone looking for a reliable comparison should be careful. I would not cite this for the SOTA claim. As a referee, I would not send it to review as is; the evaluation needs external baselines, a public dataset, and significance testing. If those were added, the method could be worth another look, but the current version's central claim does not hold.

Referee Report

3 major / 5 minor

Summary. The paper proposes MM-FusionNet, a multimodal fake-news detector that combines fine-tuned LLM text features (Vicuna-7B/Mistral-7B) with CLIP ViT-B/16 image features using a Context-Aware Dynamic Fusion Module (CADFM). CADFM consists of bidirectional cross-modal attention followed by a learned scalar gating network that adaptively weights the two modalities before concatenation and MLP classification. The model is evaluated on a private dataset called LMFND (~80,000 samples), where it reportedly achieves F1=0.938 and is claimed to be state-of-the-art, surpassing previous multimodal baselines by about 0.5%. The paper also reports ablations, a small human evaluation, gating-weight statistics, robustness to missing/noisy modalities, and a qualitative error analysis.

Significance. If the empirical claims were properly supported, the dynamic gating idea would be a simple, interpretable contribution to multimodal fake news detection. The architecture is clearly described, the equations are coherent, and the internal ablations are directionally consistent: removing gating lowers F1 from 0.938 to 0.933, and removing cross-modal attention lowers it further to 0.925. The gating-weight statistics also indicate that the model does vary modal emphasis across inputs. However, the paper's headline claim of 'state-of-the-art' performance is not established: all multimodal baselines are self-constructed, no published detector is compared against, and the sole benchmark is a private dataset with no release or detailed protocol. The work is therefore more a plausibility study of a fusion module than a demonstrated advance over existing methods. The paper also provides no code, no data, and no multi-seed statistics, further limiting its scientific value.

major comments (3)

[Abstract; Section IV-B-1; Table I] The central claim that MM-FusionNet 'surpasses existing multi-modal baselines by approximately 0.5%' is not supported by the experiments. Both multimodal baselines in Table I are constructed in this paper: Simple Concatenation Fusion is a self-built baseline, and Fixed-Weight Cross-Modal Attention is explicitly described in Section IV-C-1 as MM-FusionNet without the dynamic gating network. No published fake-news detector (e.g., SpotFake [15], MVAE, EANN, or the entity-oriented alignment method [19]) appears in Table I. The reported gain is therefore only an internal ablation gap, not a comparison with existing state-of-the-art. The authors must add external or re-implemented published baselines on the same benchmark before claiming SOTA.
[Section IV-A-1] LMFND is introduced as a 'large-scale' dataset of approximately 80,000 samples, but no public URL, release plan, or repository is provided. The paper also does not report the train/validation/test split sizes, per-split class balance, annotation guidelines, or inter-annotator agreement. Since every quantitative result in the paper depends on this private dataset, the findings cannot be independently verified or reproduced. The authors should either release LMFND with detailed documentation or repeat the key experiments on a public benchmark such as Fakeddit, Weibo, or Twitter, and report the split and annotation details.
[Tables I-IV] All performance numbers are single-run point estimates without standard deviations, confidence intervals, or significance tests. The headline advantage over Fixed-Weight Cross-Modal Attention is 0.005 in absolute F1 (0.938 vs 0.933); without repeated runs or statistical testing, this difference may be within normal training variance. The same issue affects the robustness comparisons in Table IV and the human-evaluation comparison in Table II. Reporting mean and standard deviation over multiple seeds, along with a significance test, is necessary to support wording such as 'significant gain' and 'remarkably close to human-level.'

minor comments (5)

[Section I (Contributions)] Typo: 'an novel LVLM-based framework' should be 'a novel LVLM-based framework.'
[Abstract; Section IV-B-2] The phrase 'approximately 0.5%' is ambiguous. The F1 difference between 0.933 and 0.938 is 0.5 percentage points absolute; if the intended meaning is relative improvement, that would be about 0.54%. Please state this explicitly.
[Section III-B/C] The paper repeatedly describes the approach as 'leveraging Large Vision-Language Models,' but the method actually uses a text-only LLM (Vicuna-7B/Mistral-7B) and a CLIP ViT as separate encoders with a learned fusion head. It is not a generative or unified LVLM. Consider renaming the framework or clearly stating that only the LLM and CLIP visual encoder are used, to avoid overclaiming.
[Section IV-D] The human evaluation reports point estimates on 500 samples but no confidence intervals. Given the small sample and potential annotator disagreement, the statement that MM-FusionNet's performance is 'remarkably close to human-level' is stronger than the evidence warrants.
[Section IV-F] The missing-modality experiments replace the input with a zero vector. This is a rather artificial corruption; a model can learn to detect zero patterns. Please clarify whether the zero vector is applied at the input level or after projection, and consider also evaluating with random replacement or feature dropout.

Circularity Check

1 steps flagged · score 6.0 of 10

Claimed 'surpassing existing multi-modal baselines by ~0.5%' is measured against the authors' own CADFM-minus-gating ablation, not any published detector.

self definitional [Abstract; Section IV-B-1 (Baselines); Section IV-C-1 (Ablation Study); Table I]
"Fixed-Weight Cross-Modal Attention: This advanced multi-modal baseline incorporates a bi-directional cross-modal attention mechanism, similar to the initial stage of our CADFM, to allow interaction between modalities. However, unlike MM-FusionNet, it does not employ a dynamic gating mechanism to adaptively weigh the importance of each modality... When the dynamic modal gating component is removed, effectively reducing the CADFM to only the bi-directional cross-modal attention mechanism where fusion weights are not adaptively learned, the model’s performance drops to an F1-score of 0.933 (as sh"

The abstract's central claim of 'surpassing existing multi-modal baselines by approximately 0.5%' is supported by Table I, whose best multi-modal row is 'Fixed-Weight Cross-Modal Attention'. However, Section IV-B-1 defines that baseline as the CADFM without dynamic gating, and Section IV-C-1 explicitly identifies it as MM-FusionNet with the dynamic modal gating component removed. Therefore the 0.5% F1 advantage is the internal ablation contribution of the gating network, not a comparison to any published multi-modal fake news detector. No external baseline (e.g., SpotFake, EANN, MVAE) appears in Table I. The 'state-of-the-art' claim thus reduces by construction to 'our model beats its own ablation by 0.005'.

full rationale

The only significant circularity is in the comparative SOTA claim. The paper's fixed-weight baseline is self-constructed and, by the paper's own ablation, is exactly the proposed model without dynamic gating; the claimed 0.5% gain over 'existing multi-modal baselines' is therefore an ablation gap, not an external benchmark comparison. The architectural derivation (Eqs. 1-14) is otherwise self-contained: encoders are pre-trained, fusion and gating are defined with learnable parameters, and the training objective is standard cross-entropy. There are no load-bearing self-citations or imported uniqueness theorems. Dataset availability and the representativeness of the baselines are reproducibility and correctness concerns, not additional circularity. The internal ablation itself is a genuine measurement, but the headline 'state-of-the-art' assertion is not independently substantiated.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. The free parameters listed are experimental hyperparameters set without sensitivity analysis. The central claim rests on the domain assumptions that the private LMFND dataset is valid and the self-defined baselines are a fair comparison set.

free parameters (5)

Learning rate = 1e-5
Set in Section IV-A-3 without reported sensitivity analysis.
Batch size = 32
Set in Section IV-A-3 without reported sensitivity analysis.
Training epochs = 10 (with early stopping)
Set in Section IV-A-3; early stopping criterion not quantified.
Gating dominance threshold = 0.2
Chosen by hand in Section IV-E-1 to define significant modality dominance; no justification.
Image size = 224x224
Standard choice in Section IV-A-3, not independently justified for this task.

assumptions (5)

domain assumption LMFND dataset is correctly balanced and manually annotated with a 1:1 real/fake ratio.
Introduced in Section IV-A-1; no annotation protocol, inter-annotator agreement, or dataset URL is given.
domain assumption The train/validation/test split is representative and free of leakage.
Section IV-A-3 mentions a test set but does not specify the split ratio or that images and texts are not shared between articles.
domain assumption Pre-trained encoders (Vicuna-7B or Mistral-7B, CLIP ViT-B/16) transfer effectively to the fake news detection task with fine-tuning.
Relied upon throughout the method (Section III-B); no evidence is provided for this specific domain transfer.
domain assumption The reported F1 scores are from a single run and are representative of model performance.
Section IV-B gives no standard deviation over multiple seeds or runs.
standard math Standard mathematical operations (softmax, scaled dot-product attention) are correct.
Equations (5)-(9) use standard definitions; no special assumptions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MM-FusionNet: Context-Aware Dynamic Fusion for Multi-modal Fake News Detection with Large Vision-Language Models." pith.science (2026). https://pith.science/paper/TM4PFNPY

@misc{pith2026250805684,
  author       = {Pith},
  title        = {Pith review of: MM-FusionNet: Context-Aware Dynamic Fusion for Multi-modal Fake News Detection with Large Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TM4PFNPY}},
  note         = {Machine review of arXiv:2508.05684}
}

read the original abstract

The proliferation of multi-modal fake news on social media poses a significant threat to public trust and social stability. Traditional detection methods, primarily text-based, often fall short due to the deceptive interplay between misleading text and images. While Large Vision-Language Models (LVLMs) offer promising avenues for multi-modal understanding, effectively fusing diverse modal information, especially when their importance is imbalanced or contradictory, remains a critical challenge. This paper introduces MM-FusionNet, an innovative framework leveraging LVLMs for robust multi-modal fake news detection. Our core contribution is the Context-Aware Dynamic Fusion Module (CADFM), which employs bi-directional cross-modal attention and a novel dynamic modal gating network. This mechanism adaptively learns and assigns importance weights to textual and visual features based on their contextual relevance, enabling intelligent prioritization of information. Evaluated on the large-scale Multi-modal Fake News Dataset (LMFND) comprising 80,000 samples, MM-FusionNet achieves a state-of-the-art F1-score of 0.938, surpassing existing multi-modal baselines by approximately 0.5% and significantly outperforming single-modal approaches. Further analysis demonstrates the model's dynamic weighting capabilities, its robustness to modality perturbations, and performance remarkably close to human-level, underscoring its practical efficacy and interpretability for real-world fake news detection.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 21 canonical work pages

[15]

Spotfake: A multi-modal framework for fake news detection,

S. Singhal, R. R. Shah, T. Chakraborty, P. Kumaraguru, and S. Satoh, “Spotfake: A multi-modal framework for fake news detection,” in Fifth IEEE International Conference on Multimedia Big Data, BigMM 2019, Singapore, September 11-13, 2019 . IEEE, 2019, pp. 39–47

work page 2019
[19]

Entity-oriented multi- modal alignment and fusion network for fake news detection,

P. Li, X. Sun, H. Yu, Y . Tian, F. Yao, and G. Xu, “Entity-oriented multi- modal alignment and fusion network for fake news detection,” IEEE Trans. Multim., pp. 3455–3468, 2022

work page 2022
[1]

The role of libraries in the fake news era: a survey of information scientists and library science students in greece,

P. E. Skarpa and E. Garoufallou, “The role of libraries in the fake news era: a survey of information scientists and library science students in greece,” Online Inf. Rev., pp. 1205–1224, 2022

work page 2022
[2]

The origins and consequences of public trust in government: A time series analysis,

V . A. Chanley, T. J. Rudolph, and W. M. Rahn, “The origins and consequences of public trust in government: A time series analysis,” Public opinion quarterly , 2000

work page 2000
[3]

Fake news detection on social media: A data mining perspective,

K. Shu, A. Sliva, S. Wang, J. Tang, and H. Liu, “Fake news detection on social media: A data mining perspective,” SIGKDD Explor., pp. 22–36, 2017

work page 2017
[4]

Petrou and P

M. Petrou and P. Bosdogianni, Image processing - the fundamentals . Wiley, 1999

work page 1999
[5]

MFIR: multimodal fusion and inconsistency reasoning for explainable fake news detection,

L. Wu, Y . Long, C. Gao, Z. Wang, and Y . Zhang, “MFIR: multimodal fusion and inconsistency reasoning for explainable fake news detection,” Inf. Fusion, p. 101944, 2023

work page 2023
[6]

A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly,

Y . Yao, J. Duan, K. Xu, Y . Cai, E. Sun, and Y . Zhang, “A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly,” CoRR, 2023

2023

Show all 27 references

[7]

Modeling event-pair relations in external knowledge graphs for script reasoning,

Y . Zhou, X. Geng, T. Shen, J. Pei, W. Zhang, and D. Jiang, “Modeling event-pair relations in external knowledge graphs for script reasoning,” Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 2021

2021
[8]

Claret: Pre-training a correlation-aware context-to-event transformer for event-centric gener- ation and classification,

Y . Zhou, T. Shen, X. Geng, G. Long, and D. Jiang, “Claret: Pre-training a correlation-aware context-to-event transformer for event-centric gener- ation and classification,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

2022
[9]

Eventbert: A pre- trained model for event correlation reasoning,

Y . Zhou, X. Geng, T. Shen, G. Long, and D. Jiang, “Eventbert: A pre- trained model for event correlation reasoning,” in Proceedings of the ACM Web Conference 2022 , 2022, pp. 850–859

2022
[10]

Sda: simple discrete augmentation for contrastive sentence representation learning,

D. Zhu, Z. Mao, J. Lu, R. Zhao, and F. Tan, “Sda: simple discrete augmentation for contrastive sentence representation learning,” arXiv preprint arXiv:2210.03963, 2022

2022 arXiv
[11]

Weak to strong generalization for large language models with multi-capabilities,

Y . Zhou, J. Shen, and Y . Cheng, “Weak to strong generalization for large language models with multi-capabilities,” in The Thirteenth International Conference on Learning Representations , 2025

2025
[12]

Lvlm-ehub: A comprehensive evaluation bench- mark for large vision-language models,

P. Xu, W. Shao, K. Zhang, P. Gao, S. Liu, M. Lei, F. Meng, S. Huang, Y . Qiao, and P. Luo, “Lvlm-ehub: A comprehensive evaluation bench- mark for large vision-language models,” IEEE Trans. Pattern Anal. Mach. Intell., pp. 1877–1893, 2025

2025
[13]

Visual in-context learning for large vision-language models,

Y . Zhou, X. Li, Q. Wang, and J. Shen, “Visual in-context learning for large vision-language models,” in Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024 . Association for Computational Linguis- tics...

2024
[14]

Vislinginstruct: Elevating zero-shot learning in multi-modal lan- guage models with autonomous instruction optimization,

D. Zhu, X. Tang, W. Han, J. Lu, Y . Zhao, G. Xing, J. Wang, and D. Yin, “Vislinginstruct: Elevating zero-shot learning in multi-modal lan- guage models with autonomous instruction optimization,” arXiv preprint arXiv:2402.07398, 2024

2024 arXiv
[16]

Feature importance in the age of explain- able AI: case study of detecting fake news & misinformation via a multi-modal framework,

A. Kumar and J. W. Taylor, “Feature importance in the age of explain- able AI: case study of detecting fake news & misinformation via a multi-modal framework,” Eur. J. Oper. Res., pp. 401–413, 2024

2024
[17]

Multi-modal fake news detection on social media via multi-grained information fusion,

Y . Zhou, Y . Yang, Q. Ying, Z. Qian, and X. Zhang, “Multi-modal fake news detection on social media via multi-grained information fusion,” in Proceedings of the 2023 ACM International Conference on Multimedia Retrieval, ICMR 2023, Thessaloniki, Greece, June 12-15, 2023 . ACM,...

2023
[18]

Multi-modal transformer using two-level visual features for fake news detection,

B. Wang, Y . Feng, X. Xiong, Y . Wang, and B. Qiang, “Multi-modal transformer using two-level visual features for fake news detection,” Appl. Intell., pp. 10 429–10 443, 2023

2023
[20]

Human cognition-based consistency inference networks for multi-modal fake news detection,

L. Wu, P. Liu, Y . Zhao, P. Wang, and Y . Zhang, “Human cognition-based consistency inference networks for multi-modal fake news detection,” IEEE Trans. Knowl. Data Eng. , pp. 211–225, 2024

2024
[21]

Multi-modal misinformation detection: Approaches, chal- lenges and opportunities,

S. Abdali, “Multi-modal misinformation detection: Approaches, chal- lenges and opportunities,” CoRR, 2022

2022
[22]

An effective strategy for multi-modal fake news detection,

X. Peng and X. Bao, “An effective strategy for multi-modal fake news detection,” Multim. Tools Appl., pp. 13 799–13 822, 2022

2022
[23]

Multimodal fusion and vision- language models: A survey for robot vision,

X. Han, S. Chen, Z. Fu, Z. Feng, L. Fan, D. An, C. Wang, L. Guo, W. Meng, X. Zhang, R. Xu, and S. Xu, “Multimodal fusion and vision- language models: A survey for robot vision,” CoRR, 2025

2025
[24]

Draw all your imagine: A holistic benchmark and agent framework for complex instruction-based image generation,

Y . Zhou, J. Yuan, and Q. Wang, “Draw all your imagine: A holistic benchmark and agent framework for complex instruction-based image generation,” arXiv preprint arXiv:2505.24787 , 2025

2025 arXiv
[25]

Less is more: Vision representation compression for efficient video generation with large language models,

Y . Zhou, J. Zhang, G. Chen, J. Shen, and Y . Cheng, “Less is more: Vision representation compression for efficient video generation with large language models,” 2024. 9

2024
[26]

Memorymamba: Memory- augmented state space model for defect recognition,

Q. Wang, H. Hu, and Y . Zhou, “Memorymamba: Memory- augmented state space model for defect recognition,” arXiv preprint arXiv:2405.03673, 2024

2024 arXiv
[27]

Divide-then-aggregate: An efficient tool learning method via parallel tool invocation,

D. Zhu, W. Shi, Z. Shi, Z. Ren, S. Wang, L. Yan, and D. Yin, “Divide-then-aggregate: An efficient tool learning method via parallel tool invocation,” arXiv preprint arXiv:2501.12432 , 2025. 10

2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [15]

Spotfake: A multi-modal framework for fake news detection,

S. Singhal, R. R. Shah, T. Chakraborty, P. Kumaraguru, and S. Satoh, “Spotfake: A multi-modal framework for fake news detection,” in Fifth IEEE International Conference on Multimedia Big Data, BigMM 2019, Singapore, September 11-13, 2019 . IEEE, 2019, pp. 39–47

work page 2019

[2] [19]

Entity-oriented multi- modal alignment and fusion network for fake news detection,

P. Li, X. Sun, H. Yu, Y . Tian, F. Yao, and G. Xu, “Entity-oriented multi- modal alignment and fusion network for fake news detection,” IEEE Trans. Multim., pp. 3455–3468, 2022

work page 2022

[3] [1]

The role of libraries in the fake news era: a survey of information scientists and library science students in greece,

P. E. Skarpa and E. Garoufallou, “The role of libraries in the fake news era: a survey of information scientists and library science students in greece,” Online Inf. Rev., pp. 1205–1224, 2022

work page 2022

[4] [2]

The origins and consequences of public trust in government: A time series analysis,

V . A. Chanley, T. J. Rudolph, and W. M. Rahn, “The origins and consequences of public trust in government: A time series analysis,” Public opinion quarterly , 2000

work page 2000

[5] [3]

Fake news detection on social media: A data mining perspective,

K. Shu, A. Sliva, S. Wang, J. Tang, and H. Liu, “Fake news detection on social media: A data mining perspective,” SIGKDD Explor., pp. 22–36, 2017

work page 2017

[6] [4]

Petrou and P

M. Petrou and P. Bosdogianni, Image processing - the fundamentals . Wiley, 1999

work page 1999

[7] [5]

MFIR: multimodal fusion and inconsistency reasoning for explainable fake news detection,

L. Wu, Y . Long, C. Gao, Z. Wang, and Y . Zhang, “MFIR: multimodal fusion and inconsistency reasoning for explainable fake news detection,” Inf. Fusion, p. 101944, 2023

work page 2023

[8] [6]

A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly,

Y . Yao, J. Duan, K. Xu, Y . Cai, E. Sun, and Y . Zhang, “A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly,” CoRR, 2023

2023