Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Zero-Shot Warning Generation for Misinformative Multimodal Content

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper establishes that an attention-based consistency network can detect out-of-context image–caption pairs at 87.04% accuracy and, with the same scores, prompt a frozen vision-language model to generate zero-shot contextual warnings…

desk verdict The 87.04% headline is missing from Table 2, so the paper's main claim can't be verified as submitted, though the approach is worth a referee's time. read the letter →

arxiv 2502.00752 v1 pith:DBFWUL4N submitted 2025-02-02 cs.AI cs.CLcs.IR

Giovanni Pio Delvecchio , Huy Hong Nguyen , Isao Echizen This is my paper · ORCID
classification cs.AIcs.CLcs.IR
keywords out-of-contextmisinformationmultimodalfact-checkingzero-shotwarninggenerationconsistencycheckingevidenceretrievalattentionmechanismvision-languagemodelNewsCLIPpings
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

The paper is trying to establish that out-of-context image–caption misinformation can be detected and explained in a single pipeline with modest training cost. Its attention-based consistency network compares a query pair against web-retrieved evidence along five paired signals and reaches 87.04% test accuracy (84.78% with a 5.2M-parameter variant), beating the 84.7% CCN baseline, and a frozen MiniGPT-4 then generates zero-shot contextual warnings that human raters scored 3.5/5 for informativeness and 4/5 for overall quality. This matters because most existing detectors return a score or attention weights but no debunking explanation, and explanations are what actually help readers decide what to trust.

What carries the argument

The load-bearing object is the attention-based consistency block: multi-head self-attention over retrieved evidence followed by cosine similarity, applied independently to image, label, caption, and source-page embeddings, plus a VLM-based multimodal logit for the input pair. A page-page self-attention block ranks the retrieved source pages by importance, and the top pages are included in the prompt to a frozen MiniGPT-4 along with the consistency scores and the predicted class, prompting it to write a contextual explanation or a warning. Frozen pre-trained encoders keep training cheap and modular, and the late-fusion design means any of the blocks can be dropped or swapped without retraining the whole system.

What would settle it

Take the NewsCLIPpings test set and withhold the correct source page for a random 20% of queries, replacing it with an unrelated page. If classification accuracy and human informativeness ratings do not drop, then the evidence-retrieval and page-ranking machinery is not carrying the claim; the paper's Section 5.6 failures predict a clear drop.

Watch

Extended reading notes

Core claim

The central discovery is that consistency checking over five paired signals—image-to-image, label-to-label, caption-to-caption, source-page-to-source-page, and a multimodal image–caption score—can be combined with late fusion and a linear classification head to detect out-of-context repurposing, and that the same scores can be repurposed as context for warning generation. Using frozen sentence and vision transformers plus a small trained head, the lightweight model reaches 84.78% accuracy at 5.2M parameters and trains in about three and a half hours on one GPU; the full model, which also trains CLIP, reaches 87.04% accuracy. In human evaluation, generated warnings scored 3.5/5 on informativeness and 4/5 on overall quality, with the worst failures traced to missing or noisy evidence retrieval rather than to the reasoning blocks themselves.

Load-bearing premise

The whole method assumes that for every query, the web search returns the correct source pages and evidence; when the query image or caption is missing from that evidence, the paper's own examples show the classification and the warning go wrong.

Editorial extensions

If this is right

  • Out-of-context detection no longer requires expensive backbone fine-tuning: the lightweight model trains in about 3.5 hours on one GPU and stays at 84.78% accuracy.
  • Each query pair gets a natural-language warning plus links to the top-ranked source pages, giving fact-checkers and readers something they can inspect rather than just a score.
  • The page-page attention block identifies which retrieved source is most relevant for the decision, so classification and the explanation are driven by the same evidence.
  • The zero-shot prompt design means the warning generator can be replaced by any instruction-following VLM without retraining, as long as the consistency scores and source pages are supplied.
  • The reported accuracy positions the method between ESCNet and SNIFFER while adding explanation generation that those systems lack or obtain only through additional Q-Former fine-tuning.

Reading between the lines

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

  • An extension the paper leaves implicit: the consistency scores could gate the warning generator, suppressing or downgrading explanations when the evidence is weak; the paper's own qualitative failures suggest this would improve reliability.
  • Because the method leans on inverse image search, its accuracy outside news-style domains—where the same image rarely appears in a captioned article—is likely to degrade, a testable extension on non-news datasets.
  • The human evaluation's 3.5/5 informativeness score probably overstates performance on falsified pairs, because pristine pairs are easier to contextualize; a separate evaluation restricted to falsified pairs would be a sharper test.
  • The page-page attention block is a generic evidence reranker and could be lifted into other retrieval-augmented multimodal tasks, though the paper does not claim this.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes a pipeline for detecting out-of-context (OOC) multimodal misinformation in the NewsCLIPpings benchmark. The pipeline retrieves external evidence (images, captions, and source pages), computes five consistency scores through attention-based blocks, and feeds these scores with selected source pages into a frozen MiniGPT-4 to generate zero-shot contextual warnings or explanations. The authors report 87.04% test accuracy for a full-scale CLIP-based model, 84.78% for a lightweight MiniGPT-4-based model with 5.2M parameters, and human-evaluation scores of 3.5/5 for informativeness and 4/5 for overall quality of the generated warnings. The paper positions these results against prior work (CCN, ESCNet, SNIFFER) and claims reduced training time and the addition of an explanation capability.

Significance. If the reported accuracies and human-evaluation scores are reliable, the paper would make a useful contribution by coupling OOC detection with explanation generation at low training cost. The lightweight 5.2M-parameter model and the use of source-page ranking are interesting design choices, and the comparison against several baselines is a strength. However, the headline accuracy is currently unverifiable from Table 2, and the human evaluation is too weakly reported to support the warning-quality claims. The dependence on an unreleased evidence-retrieval pipeline further limits reproducibility. The core idea is defensible, but the evidence presented in the manuscript needs substantial strengthening before the central claims can be accepted.

major comments (3)
  1. [§5.2, Table 2, Abstract] The paper's headline accuracy of 87.04% is not reported in Table 2: the highest CLIP-backed test accuracy shown is 86.70% (model 10), and no row corresponds to a configuration trained with the rescaled learning rate. Section 5.2 says the full model reaches 87.04% with the rescaled rate and 86.70% with the standard rate, but no experimental setup, seed, or run is specified for the former. Because the claimed advantage over the 84.7% CCN baseline rests on this number, the central detection claim cannot currently be checked. The table also lists single-run test accuracies for 24 variants with no variance or confidence intervals; selecting the best variant on the same test set can inflate the reported edge.
  2. [§5.5, §4.6] The human evaluation of warning generation is too weakly specified to support the claim that the warnings are informative and high quality. Only 100 test samples were used, each of the 20 raters assessed just 5 samples, and the paper reports only mean Informativeness (3.5) and Overall Quality (4.0) with no standard deviations, per-item distributions, or inter-annotator agreement. Because the prompt in §4.6 already inserts the classifier's score and binary label, the raters may be evaluating a rationalization of the system's own prediction rather than an independent explanation; at minimum, the evaluation should report rater instructions, randomization, and agreement.
  3. [§3.2, §5.6] The method's inputs are the evidence and source pages retrieved by the pipeline of Abdelnabi et al. [1], but neither that evidence nor the re-downloaded source pages are released, making the experiments non-reproducible. Section 5.6 also documents failure cases where the query caption or query image is missing from the retrieved evidence, leading to incorrect warnings and one misclassification. The paper should release the evidence URLs/source pages or clearly state their availability, and should quantify how often the query item is absent from the evidence, since this directly conditions the reported accuracy and warning quality.
minor comments (5)
  1. [Table 2] Table 2 uses commas as decimal separators (e.g., 86,46), which is confusing and inconsistent with the rest of the text; use decimal points throughout.
  2. [§5.2] The phrase "highlighted in green" in §5.2 cannot be verified in the text version; ensure the relevant table rows are clearly marked in the published form.
  3. [§4.5] In the loss formula, the notation ln is used both for the log function and for the indexed loss terms; clarify the notation to avoid confusion.
  4. [§5.6] The description of the third sample says the prediction was correct but the explanation was inconsistent; this distinction between prediction correctness and explanation correctness should be stated explicitly rather than implied.
  5. [§4.6] The term "zero-shot" is appropriate in the sense that MiniGPT-4 is not fine-tuned, but because the prompt includes the classifier's score and binary label, the generation is classifier-conditioned; consider describing it as such to avoid overclaiming.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported accuracy advantage is partly a test-set selection artifact: the lightweight model is chosen as the best of 12 variants by test accuracy and that same selected maximum is then reported as the model's performance.

  1. fitted input called prediction [Section 5.2 and 5.3, Table 2]
    "Our lightweight model, employing frozen MiniGPT-4 as VLM, achieves 84.8% of accuracy with 5.2 million parameters, it is trainable in 3 hours and 38 minutes on a single GPU, while it only requires 13 minutes on three GPUs. It is highlighted in blue in table 2. ... Nevertheless, we consider version 17 to be the best model version using the MiniGPT-4 model as it had the highest accuracy and fewer parameters while using all our attention blocks."

    Version 17 is selected as 'best' because it 'had the highest accuracy' in Table 2, and Table 2's accuracy column is the test accuracy. The reported 84.78% is therefore the maximum of the MiniGPT-4 test-accuracy rows, not an independent holdout estimate. Choosing a model by the same test-set accuracy that is then reported as the achievement makes the headline number a selection statistic: by construction it equals max over variants, so the 0.08-point margin over CCN (84.7) is an artifact of test-set selection. The full-model 87.04% claim does not appear in Table 2, so the main accuracy comparison rests on an unverifiable selected maximum.

full rationale

The core architecture (attention-based consistency blocks, late fusion, classification head) is trained and evaluated on the NewsCLIPpings split and does not itself reduce to its inputs. The warning generator is a frozen MiniGPT-4 prompted with the classifier score and retrieved evidence; the paper discloses that the verdict is inserted into the prompt, so the generated warning's label is inherited from the classifier rather than independently derived. That is a dependence, not a circular derivation. The one load-bearing circular step is the model-selection procedure: the lightweight variant (and implicitly the full variant) is chosen as the best of 24 configurations by test accuracy, then that same selected test accuracy is reported as the model's performance. The reported numbers are thus maxima over configurations rather than predictive estimates, which materially inflates the comparison against the 84.7% CCN baseline. Self-citations (Shalabi et al. [24]) are baselines and not load-bearing. Score 6 reflects one prediction that reduces by construction while the remainder of the pipeline has independent content.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the external evidence pipeline and on embedding-space cosine similarity as a proxy for consistency; these are domain assumptions, not derived results. No new entities are introduced. The main free parameters are standard training hyperparameters plus the threshold and prompt-construction choices, none of which are subjected to sensitivity analysis.

free parameters (5)
  • classification threshold th = 0.5 default or th_EER from validation (e.g., 0.5247)
    Decision boundary is chosen using validation EER; test accuracy changes between th=0.5 and th_EER in Table 2.
  • multimodal block hidden dimension = 256
    The linear layer inside the image-caption consistency block projects to 256 dimensions; no sensitivity analysis is provided.
  • top source pages in prompt = 4 pages
    The warning prompt includes the 4 highest-attention pages; the number is chosen by hand and not ablated.
  • source page snippet length = 400 characters
    Only the first 400 characters of each selected page are included in the prompt; no variation is tested.
  • learning rates and schedule = 9e-5 and 5e-4 with cyclic schedule
    These training hyperparameters affect the reported accuracy and are chosen by the authors without a full search.
assumptions (4)
  • domain assumption The NewsCLIPpings merge-balanced subset is a valid benchmark for real-world out-of-context misinformation detection.
    All accuracy claims are measured on this synthetic dataset; real-world OOC distribution may differ (Section 5.1).
  • domain assumption The external evidence collection from Abdelnabi et al. is complete and correctly associated with each query image/caption.
    The pipeline reuses evidence and URLs from [1] (Section 3.2). Failure cases in Section 5.6 show that missing evidence directly harms performance.
  • domain assumption Cosine similarity in frozen embedding spaces is a meaningful proxy for semantic and visual consistency.
    All consistency scores are cosine similarities over ViT, Sentence-BERT, and CLIP/MiniGPT-4 embeddings (Sections 4.1 to 4.4).
  • domain assumption The binary labels in NewsCLIPpings are complete and correct.
    Training and evaluation use the dataset labels as ground truth without auditing them (Section 3.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Zero-Shot Warning Generation for Misinformative Multimodal Content." pith.science (2026). https://pith.science/paper/DBFWUL4N

@misc{pith2026250200752,
  author       = {Pith},
  title        = {Pith review of: Zero-Shot Warning Generation for Misinformative Multimodal Content},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DBFWUL4N}},
  note         = {Machine review of arXiv:2502.00752}
}
read the original abstract

The widespread prevalence of misinformation poses significant societal concerns. Out-of-context misinformation, where authentic images are paired with false text, is particularly deceptive and easily misleads audiences. Most existing detection methods primarily evaluate image-text consistency but often lack sufficient explanations, which are essential for effectively debunking misinformation. We present a model that detects multimodal misinformation through cross-modality consistency checks, requiring minimal training time. Additionally, we propose a lightweight model that achieves competitive performance using only one-third of the parameters. We also introduce a dual-purpose zero-shot learning task for generating contextualized warnings, enabling automated debunking and enhancing user comprehension. Qualitative and human evaluations of the generated warnings highlight both the potential and limitations of our approach.

Figures

Figures reproduced from arXiv: 2502.00752 by the authors.

Figure 1
Figure 1. Proposed pipeline for debunking misleading content. Each input image [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed architecture for multimodal misinformation detection: Each consistency block provides a score in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 27 canonical work pages

  1. [1]

    Open- domain, content-based, multi-modal fact-checking of out-of- context images via online resources

    Sahar Abdelnabi, Rakibul Hasan, and Mario Fritz. Open- domain, content-based, multi-modal fact-checking of out-of- context images via online resources. InCVPR, pages 14940– 14949, 2022. 2, 5, 6

    work page 2022

  2. [2]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 ,

    arXiv

  3. [3]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in NeurIPS, 35:23716–23736, 2022. 4

    work page 2022

  4. [4]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. Ieee, 2009. 3

    work page 2009

  5. [5]

    Early vs late fusion in multimodal convolutional neural networks

    Konrad Gadzicki, Razieh Khamsehashari, and Christoph Zetzsche. Early vs late fusion in multimodal convolutional neural networks. In FUSION, pages 1–6. IEEE, 2020. 3

    work page 2020

  6. [6]

    From im- ages to textual prompts: Zero-shot visual question answering with frozen large language models

    Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From im- ages to textual prompts: Zero-shot visual question answering with frozen large language models. In CVPR, pages 10867– 10877, 2023. 4

    work page 2023

  7. [7]

    A picture paints a thousand lies? the effects and mechanisms of multimodal disinformation and rebuttals disseminated via social media

    Michael Hameleers, Thomas E Powell, Toni GLA Van Der Meer, and Lieke Bos. A picture paints a thousand lies? the effects and mechanisms of multimodal disinformation and rebuttals disseminated via social media. Political com- munication, 37(2):281–301, 2020. 1

    work page 2020

  8. [8]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016. 3

    2016

Show all 32 references
  1. [9]

    Batch normalization: Accelerating deep network training by reducing internal co- variate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal co- variate shift. In ICML, pages 448–456. PMLR, 2015. 3, 4

    2015

  2. [10]

    Rlaif: Scaling reinforcement learning from human feedback with ai feedback

    Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop, Victor Carbune, and Abhinav Rastogi. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023. 4

    2023 arXiv

  3. [11]

    Visual news: Benchmark and challenges in news im- age captioning

    Fuxiao Liu, Yinghan Wang, Tianlu Wang, and Vicente Or- donez. Visual news: Benchmark and challenges in news im- age captioning. In EMNLP, pages 6761–6771, 2021. 2

    2021

  4. [12]

    Robust domain misinformation detection via multi- modal feature alignment

    Hui Liu, Wenya Wang, Hao Sun, Anderson Rocha, and Hao- liang Li. Robust domain misinformation detection via multi- modal feature alignment. IEEE T-IFS, 2023. 2

    2023

  5. [13]

    Newsclip- pings: Automatic generation of out-of-context multimodal media

    Grace Luo, Trevor Darrell, and Anna Rohrbach. Newsclip- pings: Automatic generation of out-of-context multimodal media. In EMNLP, pages 6801–6817, 2021. 1, 2, 5, 6

    2021

  6. [14]

    fake news

    Tarlach McGonagle. “fake news” false fears or real con- cerns? Netherlands Quarterly of Human Rights, 35(4):203– 209, 2017. 1

    2017

  7. [15]

    Representation biases in sentence transformers

    Dmitry Nikolaev and Sebastian Pad ´o. Representation biases in sentence transformers. In EACL, pages 3701–3716, 2023. 3

    2023

  8. [16]

    Dinov2: Learning robust visual features without su- pervision

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without su- pervision. TMLR, 2023. 3, 5

    2023

  9. [17]

    Training lan- guage models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Car- roll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training lan- guage models to follow instructions with human feedback. Advances in NeurIPS, 35:27730–27744, 2022. 4

    2022

  10. [18]

    A com- parative analysis of early and late fusion for the multimodal two-class problem

    Luis M Pereira, Addisson Salazar, and Luis Vergara. A com- parative analysis of early and late fusion for the multimodal two-class problem. IEEE Access, 2023. 3

    2023

  11. [19]

    Declare: Debunking fake news and false claims using evidence-aware deep learning

    Kashyap Popat, Subhabrata Mukherjee, Andrew Yates, and Gerhard Weikum. Declare: Debunking fake news and false claims using evidence-aware deep learning. In EMNLP, pages 22–32, 2018. 2

    2018

  12. [20]

    Sniffer: Multimodal large language model for explainable out-of-context misinformation detection

    Peng Qi, Zehong Yan, Wynne Hsu, and Mong Li Lee. Sniffer: Multimodal large language model for explainable out-of-context misinformation detection. In CVPR, pages 13052–13062, 2024. 2, 5, 6

    2024

  13. [21]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In ICML, pages 8748–8763. PMLR, 2021. 3, 5

    2021

  14. [22]

    Do vision trans- formers see like convolutional neural networks? Advances in NeurIPS, 34:12116–12128, 2021

    Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision trans- formers see like convolutional neural networks? Advances in NeurIPS, 34:12116–12128, 2021. 3

    2021

  15. [23]

    Sentence-bert: Sen- tence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sen- tence embeddings using siamese bert-networks. In EMNLP- IJCNLP, pages 3982–3992, 2019. 3, 5

    2019

  16. [24]

    Leveraging chat-based large vision language mod- els for multimodal out-of-context detection

    Fatma Shalabi, Hichem Felouat, Huy H Nguyen, and Isao Echizen. Leveraging chat-based large vision language mod- els for multimodal out-of-context detection. In AINA, 2024. 2, 5, 6

    2024

  17. [25]

    Human evaluation of automatically gener- ated text: Current trends and best practice guidelines

    Chris van der Lee, Albert Gatt, Emiel van Miltenburg, and Emiel Krahmer. Human evaluation of automatically gener- ated text: Current trends and best practice guidelines. Com- puter Speech & Language, 67:101151, 2021. 7

    2021

  18. [26]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. NIPS, 30, 2017. 3, 7

    2017

  19. [27]

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers

    Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Ad- vances in NeurIPS, 33:5776–5788, 2020. 3, 5

    2020

  20. [28]

    Visual transform- ers: Token-based image representation and processing for computer vision

    Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual transform- ers: Token-based image representation and processing for computer vision. arXiv preprint arXiv:2006.03677 , 2020. 3, 5

    2006 arXiv

  21. [29]

    End-to-end multimodal fact-checking and explanation generation: A challenging dataset and mod- els

    Barry Menglong Yao, Aditya Shah, Lichao Sun, Jin-Hee Cho, and Lifu Huang. End-to-end multimodal fact-checking and explanation generation: A challenging dataset and mod- els. In ACM SIGIR, pages 2733–2743, 2023. 2, 5, 6, 7 9

    2023

  22. [30]

    Escnet: Entity-enhanced and stance checking network for multi-modal fact-checking

    Fanrui Zhang, Jiawei Liu, Jingyi Xie, Qiang Zhang, Yongchao Xu, and Zheng-Jun Zha. Escnet: Entity-enhanced and stance checking network for multi-modal fact-checking. In WWW, pages 2429–2440, 2024. 2, 5, 6

    2024

  23. [31]

    Detecting out-of-context multimodal misinformation with interpretable neural-symbolic model

    Yizhou Zhang, Loc Trinh, Defu Cao, Zijun Cui, and Yan Liu. Detecting out-of-context multimodal misinformation with interpretable neural-symbolic model. arXiv preprint arXiv:2304.07633, 2023. 2

    2023 arXiv

  24. [32]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mo- hamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2023. 2, 3, 5 10

    2023

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.