Pith. sign in

REVIEW 4 major objections 5 minor 47 references

E2LVLM:Evidence-Enhanced Large Vision-Language Model for Multimodal Out-of-Context Misinformation Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that reranking and rewriting search-retrieved evidence before feeding it to a vision-language model yields state-of-the-art out-of-context misinformation detection, with 89.9% overall accuracy and 90.3% on falsified…

desk verdict Useful evidence-processing recipe for LVLM fact-checking, but the SOTA claim needs error bars and the self-generated explanation supervision is unvalidated. read the letter →

arxiv 2502.10455 v1 pith:NLYSNF6N submitted 2025-02-12 cs.LG cs.MM

Junjie Wu , Yumeng Fu , Nan Yu , Guohong Fu This is my paper · ORCID
classification cs.LGcs.MM
keywords out-of-contextmisinformationmultimodalfact-checkingvision-languagemodelsevidencererankingrewritinginstructiontuningexplainabilityNewsCLIPpings
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 is trying to establish that the way textual evidence is prepared before being fed to a large vision-language model matters as much as the model itself for detecting out-of-context image-text pairs. E2LVLM reranks search-retrieved evidence to pick the most relevant item, rewrites it into coherent natural language aligned with the model's inputs, and then fine-tunes the model on an instruction-following dataset that includes both judgments and explanations. The authors show that this pipeline reaches 89.9% overall accuracy on the NewsCLIPpings benchmark and 90.3% on falsified samples, outperforming prior methods including SNIFFER. They also report 74.4% accuracy on the VERITE True vs OOC split. If correct, the work suggests that cleaning and aligning external evidence is a cheap but effective route to better multimodal misinformation detection and explainability.

What carries the argument

The central mechanism is a two-level evidence adaptation pipeline: a reranking stage in which Qwen2-VL selects the most relevant textual evidence from a set retrieved by inverse image search, and a rewriting stage in which the same model rewrites that snippet into natural-language content that is closer to the image's semantics and the LVLM's input format. A second pillar is the automatically constructed OOC instruction-following dataset, generated by prompting Qwen2-VL to produce compelling explanations for each image-claim pair, which is then used for one-stage multimodal instruction tuning with LoRA and a next-token prediction loss. The claim is that this pipeline—not any single component alone—aligns external evidence with the model and thereby improves both detection accuracy and the quality of rationales.

What would settle it

Take a random sample of 200 test pairs from NewsCLIPpings, replace the model's rewritten evidence with independently verified evidence from the original source articles, and compare detection accuracy; if accuracy does not drop or even improves, the rewriting step's contribution to correctness is not the mechanism claimed. Alternatively, have human annotators judge whether the explanations that accompany correct judgments are factually consistent with the source articles; if a substantial share of correct judgments rest on hallucinated rationales, the claimed link between explanation quality and detection performance would be broken.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that retrieved textual evidence about an image is too noisy and too far from the model's input distribution to help vision-language models distinguish falsified from pristine image-claim pairs, so evidence should be adapted in two stages: an LVLM-based reranking step selects the single most relevant evidence snippet, and a rewriting step converts it into coherent, contextually attuned content. With this adapted evidence, plus an instruction-tuning dataset that supplies both a judgment and a plausible explanation for each pair, E2LVLM fine-tunes Qwen2-VL-7B and achieves 89.9% accuracy overall, 90.3% on falsified samples, and 89.4% on pristine samples on NewsCLIPpings, exceeding the previous state of the art by about 1.5 points overall and 3.4 points on falsified samples.

Load-bearing premise

The rewritten evidence and the explanation annotations used as training supervision are both generated by the same Qwen2-VL model family that is later evaluated, and the paper assumes these self-generated texts are factually faithful, with no human or external consistency check.

Editorial extensions

If this is right

  • E2LVLM outperforms the prior state of the art, SNIFFER, by about 1.5 points overall and 3.4 points on falsified samples on the NewsCLIPpings Merged/Balance subset.
  • On VERITE's True vs OOC split, E2LVLM achieves 74.4% accuracy without any additional training on that dataset, outperforming RED-DOT by 0.5 points and SNIFFER by 0.4 points.
  • Each component of the pipeline contributes: textual evidence use raises zero-shot accuracy from 69.1% to 76.7%, reranking adds 9.4 points on falsified samples, rewriting adds further gains, and explanation supervision lifts falsified accuracy to 90.3%.
  • Instruction tuning with the constructed dataset lets a 7.6B-parameter model reach 89.9% overall accuracy, while increasing to 72B parameters yields only 0.7 additional points, suggesting a favorable compute-accuracy trade-off.
  • E2LVLM can be deployed with only 10% of the training data and still beat several earlier methods, supporting early-stage misinformation detection.

Reading between the lines

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

  • Because the evidence-adaptation steps use the same LVLM that is later fine-tuned, the pipeline may be vulnerable to self-confirmation: if the base model already believes an image-claim pair is falsified, it could rewrite evidence to fit that belief, which would inflate benchmark accuracy without improving real-world factuality.
  • The rerank-and-rewrite recipe is task-agnostic and could be applied to other multimodal verification tasks, such as deepfake claims or propaganda analysis, where retrieved evidence also needs to be aligned with the model's input space.
  • A direct testable extension would be to use a different, smaller model to generate the rewritten evidence and explanations during training, then evaluate whether the final detector still gains; if the gains vanish, the benefit comes from the teacher's knowledge rather than from evidence alignment.
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

4 major / 5 minor

Summary. The paper proposes E2LVLM, an evidence-enhanced large vision-language model for multimodal out-of-context (OOC) misinformation detection. The method first uses Qwen2-VL to rerank and rewrite textual evidence retrieved by an inverse image search, then constructs a multimodal instruction-following dataset whose explanations are themselves generated by Qwen2-VL, and finally fine-tunes Qwen2-VL with LoRA on this dataset to produce both judgments and rationales. On NewsCLIPpings the paper reports 89.9% overall accuracy and 90.3% accuracy on falsified samples, and on VERITE it reports 74.4% 'True vs OOC' accuracy, claiming state-of-the-art performance over SNIFFER and other baselines. The paper also includes ablations, model-size comparisons, data-scaling experiments, and qualitative distribution analyses.

Significance. If the results are robust, the paper makes a useful practical contribution: it shows that reranking and rewriting retrieved textual evidence can improve LVLM-based OOC detection, and it provides an open-source, instruction-tuned LVLM that outputs explanations alongside judgments. The detection labels come from external benchmarks, so the core accuracy claim is not circular with respect to the generated explanations. The ablation in Table 2 demonstrates that each proposed component adds accuracy, and the comparisons to prior methods follow the standard protocol used in this literature. However, the lack of statistical significance testing, a direct numeric inconsistency between the abstract and Table 1, and the unvalidated use of the same model family to generate training supervision all currently limit how strongly the state-of-the-art claim can be asserted.

major comments (4)
  1. [§4.1, Table 1, Table 5] The paper states that all results are averages of three runs but reports no standard deviations, confidence intervals, or significance tests. The improvements over SNIFFER are 1.5 points on NewsCLIPpings 'All' (89.9 vs 88.4), 3.4 points on 'Falsified' (90.3 vs 86.9), and only 0.4 points on VERITE 'True vs OOC' (74.4 vs 74.0). With three runs, gaps of this size, especially the 0.4-point VERITE gap on a 1,000-sample test set, may be within run-to-run noise. Please provide per-run results, error bars in all figures and tables, and a paired significance test (e.g., bootstrap or McNemar) for the main comparisons.
  2. [Abstract and §1 vs Table 1] The abstract and Section 1 state that E2LVLM achieves 90.34% accuracy and outperforms the state of the art by around 3.44%. Table 1, however, reports 89.9% for 'All' and 90.3% for 'Falsified'. The 3.44-point gain matches only the Falsified column (90.3 vs 86.9), not the overall accuracy. This inconsistency must be resolved, because the reader cannot tell which number is the headline result.
  3. [§3.3, §3.4, Table 4] The rewritten evidence and the explanation annotations used as supervised training signals are both generated by Qwen2-VL, the same model family that is later fine-tuned and evaluated. The paper asserts that these generated rationales are 'compelling' without any human evaluation or factual consistency check. This does not invalidate the detection-accuracy claim, since the labels come from NewsCLIPpings and VERITE, but it leaves the explainability claim unsupported and raises the risk that the model is trained to imitate its own hallucinations. Please add human evaluation of explanation quality or automated factuality metrics, and discuss the potential contamination from using the same model family for data construction and final evaluation.
  4. [Table 2 vs Table 3] There is a direct numerical conflict between the zero-shot baselines in the two tables. Table 2's first row reports raw Qwen2-VL with no evidence and no tuning at 69.1% 'All', while Table 3 reports Qwen2-VL-7B without tuning at 79.1% 'All'. Since Table 2 is the central ablation supporting the claim that each component contributes, please clarify the exact input configuration used in each table and rerun the ablations under a single consistent protocol.
minor comments (5)
  1. [References, [23]] The reference for SAFE, [23], is 'SAFE: Self-attentive function embeddings for binary similarity,' which appears to be a different paper from the SAFE baseline used in Table 1 for fake-news detection. Please verify and correct this citation.
  2. [Table 5] The 'All' column in Table 5 reports 89.9 for E2LVLM, which is the NewsCLIPpings result, but the caption says the comparison is on VERITE. The caption should state explicitly which numbers come from which dataset.
  3. [§3.3] The text says tr_e is 'the most relevant item related to the claim t' selected by cosine similarity 'unless otherwise specified.' Please clarify how visual evidence is reranked and why cosine similarity is used for visual evidence while LVLM-based reranking is used for textual evidence.
  4. [Figure 6] Figure 6 plots accuracy against training data proportions but includes no error bars, even though the paper states that all results are averaged over three runs. Add error bars for consistency.
  5. [§2.2, Figure 6 caption, reference [45]] There are several typos: 'LLaV A' in Section 2.2 should be 'LLaVA'; the Figure 6 caption writes 'E2LVM' instead of 'E2LVLM'; and reference [45] spells 'mutli-modal' instead of 'multi-modal.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; detection accuracy is anchored to external benchmark labels.

full rationale

E2LVLM's central claim is state-of-the-art accuracy on NewsCLIPpings and VERITE. Those labels are external benchmark ground truths, not outputs of the proposed pipeline, so the main result is not defined in terms of its own inputs. The evidence reranking/rewriting and explanation generation use the same Qwen2-VL family, but the ablations in Table 2 measure downstream detection accuracy on held-out samples, so the contribution of each module is empirically quantified rather than imposed by construction. The explanation annotations are model-generated training targets, but they are conditioned on supplied ground-truth labels ('generating compelling explanations to support their judgments for the falsified information'), and the reported accuracy numbers do not depend on the rationales being externally verified. The lack of significance testing and the 90.34% versus 89.9% reporting inconsistency are correctness or reporting concerns, not circularity. No load-bearing self-citation chain or imported uniqueness theorem is present; the authors do not rely on their own prior work to force the design. Hence no demonstrated circular step is found.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central detection claim rests on external benchmark labels, which are independent. The main assumptions are the sufficiency of web-retrieved evidence, the reliability of Qwen2-VL representations for measuring image-text closeness, and the quality of Qwen2-VL-generated explanations as supervision. The latter is the most fragile because it is unverified.

free parameters (3)
  • top-1 evidence count k = 1
    Selected because accuracy degrades with more evidence (Figure 4). Chosen by validation performance, not a model parameter.
  • LoRA rank and alpha = not reported
    Implementation uses LoRA but does not specify rank or alpha; these are hyperparameters chosen by the authors.
  • prompt templates = hand-crafted
    The four prompts (Prerank, Prewrite, PExpla, POOC) are manually designed; wording affects outputs and is not systematically optimized or ablated.
assumptions (3)
  • domain assumption Google API inverse search provides sufficient relevant textual and visual evidence for image-claim pairs.
    Assumed throughout; evidence gathered following CCN [1].
  • domain assumption Qwen2-VL's representations are reliable for measuring closeness between images and text (Figures 1 and 5).
    Used to motivate reranking and rewriting and to visualize distributions.
  • ad hoc to paper Explanations generated by Qwen2-VL are 'compelling' and suitable as supervised signals.
    No human evaluation or external quality metric; central to the instruction-tuning claim (Section 3.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of E2LVLM:Evidence-Enhanced Large Vision-Language Model for Multimodal Out-of-Context Misinformation Detection." pith.science (2026). https://pith.science/paper/NLYSNF6N

@misc{pith2026250210455,
  author       = {Pith},
  title        = {Pith review of: E2LVLM:Evidence-Enhanced Large Vision-Language Model for Multimodal Out-of-Context Misinformation Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NLYSNF6N}},
  note         = {Machine review of arXiv:2502.10455}
}
read the original abstract

Recent studies in Large Vision-Language Models (LVLMs) have demonstrated impressive advancements in multimodal Out-of-Context (OOC) misinformation detection, discerning whether an authentic image is wrongly used in a claim. Despite their success, the textual evidence of authentic images retrieved from the inverse search is directly transmitted to LVLMs, leading to inaccurate or false information in the decision-making phase. To this end, we present E2LVLM, a novel evidence-enhanced large vision-language model by adapting textual evidence in two levels. First, motivated by the fact that textual evidence provided by external tools struggles to align with LVLMs inputs, we devise a reranking and rewriting strategy for generating coherent and contextually attuned content, thereby driving the aligned and effective behavior of LVLMs pertinent to authentic images. Second, to address the scarcity of news domain datasets with both judgment and explanation, we generate a novel OOC multimodal instruction-following dataset by prompting LVLMs with informative content to acquire plausible explanations. Further, we develop a multimodal instruction-tuning strategy with convincing explanations for beyond detection. This scheme contributes to E2LVLM for multimodal OOC misinformation detection and explanation. A multitude of experiments demonstrate that E2LVLM achieves superior performance than state-of-the-art methods, and also provides compelling rationales for judgments.

Figures

Figures reproduced from arXiv: 2502.10455 by the authors.

Figure 1
Figure 1. Subfigures (a) and (b) show the representation dis [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of E2LVLM - the evidence-enhanced large vision-language model. Given an authentic image input together with a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Prompts and their examples in E2LVLM. (a) Reranking prompt [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Visualization of various data distributions. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Performance of E2LVM on NewsCLIPpings [22] using different training data proportions [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 35 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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14940–14949, 2022. 1, 2, 3, 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]

    Cos- mos: Catching out-of-context image misuse using self- supervised learning

    Shivangi Aneja, Chris Bregler, and Matthias Nießner. Cos- mos: Catching out-of-context image misuse using self- supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14084–14092, 2023. 1

    work page 2023

  4. [4]

    The making of an ai news anchor—and its implications

    Matyas Bohacek and Hany Farid. The making of an ai news anchor—and its implications. Proceedings of the National Academy of Sciences of the United States of America , 121 (1):e2315678121, 2024. 1

    work page 2024

  5. [5]

    Lion: Empowering multimodal large language model with dual-level visual knowledge

    Gongwei Chen, Leyang Shen, Rui Shao, Xiang Deng, and Liqiang Nie. Lion: Empowering multimodal large language model with dual-level visual knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26540–26550, 2024. 3

    work page 2024

  6. [6]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang- hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong- hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, 2023. 3, 6

    work page 2023

  7. [7]

    Instructblip: Towards general- purpose vision-language models with instruction tuning

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general- purpose vision-language models with instruction tuning. In International Conference on Neural Information Processing Systems, pages 49250–49267, 2023. 2, 3, 5

    work page 2023

  8. [8]

    Overview of the grand challenge on detect- ing cheapfakes at acm icmr 2024

    Duc-Tien Dang-Nguyen, Sohail Ahmed Khan, Michael Riegler, P˚al Halvorsen, Anh-Duy Tran, Minh-Son Dao, and Minh-Triet Tran. Overview of the grand challenge on detect- ing cheapfakes at acm icmr 2024. In International Confer- ence on Multimedia Retrieval, pages 1275–1281, 2024. 1

    work page 2024

Show all 47 references
  1. [9]

    Flashattention-2: Faster attention with better par- allelism and work partitioning

    Tri Dao. Flashattention-2: Faster attention with better par- allelism and work partitioning. In International Conference on Learning Representations, 2024. 6

    2024

  2. [10]

    Learning domain-invariant features for out- of-context news detection.arXiv preprint arXiv:2406.07430,

    Yimeng Gu, Mengqi Zhang, Ignacio Castro, Shu Wu, and Gareth Tyson. Learning domain-invariant features for out- of-context news detection.arXiv preprint arXiv:2406.07430,

    arXiv

  3. [11]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and other. Lora: Low-rank adaptation of large language models. In Interna- tional Conference on Learning Representations, 2022. 5

    2022

  4. [12]

    Gpt-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perel- man, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli- hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2

    2024 arXiv

  5. [13]

    Hallucination augmented contrastive learn- ing for multimodal large language model

    Chaoya Jiang, Haiyang Xu, Mengfan Dong, Jiaxing Chen, Wei Ye, Ming Yan, Qinghao Ye, Ji Zhang, Fei Huang, and Shikun Zhang. Hallucination augmented contrastive learn- ing for multimodal large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patte...

    2024

  6. [14]

    Llava-vsd: Large language-and-vision as- sistant for visual spatial description

    Yizhang Jin, Jian Li, Jiangning Zhang, Jianlong Hu, Zhenye Gan, Xin Tan, Yong Liu, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Llava-vsd: Large language-and-vision as- sistant for visual spatial description. In Proceedings of the 32nd ACM International Conference on Multimedi...

    2024

  7. [15]

    Geochat: Grounded large vision-language model for remote sensing

    Kartik Kuckreja, Muhammad Sohail Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad Shahbaz Khan. Geochat: Grounded large vision-language model for remote sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27831– 2784...

    2024

  8. [16]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In In- ternational Conference on Machine Learning, pages 19730– 19742, 2023. 4

    2023

  9. [17]

    Visualbert: A simple and perfor- mant baseline for vision and language

    Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perfor- mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 5, 6

    1908 arXiv

  10. [18]

    Detecting multimedia generated by large ai models: A survey

    Li Lin, Neeraj Gupta, Yue Zhang, Hainan Ren, Chun-Hao Liu, Feng Ding, Xin Wang, Xin Li, Luisa Verdoliva, and Shu Hu. Detecting multimedia generated by large ai models: A survey. arXiv preprint arXiv:2402.00045, 2024. 1

    2024 arXiv

  11. [19]

    Forgery-aware adaptive transformer for generalizable synthetic image detection

    Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Jingdong Wang, and Yao Zhao. Forgery-aware adaptive transformer for generalizable synthetic image detection. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 10770–10780, 2024. 1, 3

    2024

  12. [20]

    Fka-owl: Advancing multimodal fake news detection through knowledge-augmented lvlms

    Xuannan Liu, Peipei Li, Huaibo Huang, Zekun Li, Xing Cui, Jiahao Liang, Lixiong Qin, Weihong Deng, and Zhaofeng He. Fka-owl: Advancing multimodal fake news detection through knowledge-augmented lvlms. In Proceedings of the 32nd ACM International Conference on Multimedia , page...

    2024

  13. [21]

    Decoupled weight decay regularization

    I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6

    2017 arXiv

  14. [22]

    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 Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP 2021) , pages 6801–6817, 2021. 1, 2, 5, 6, 8

    2021

  15. [23]

    Safe: Self- attentive function embeddings for binary similarity

    Luca Massarelli, Giuseppe Antonio Di Luna, Fabio Petroni, Roberto Baldoni, and Leonardo Querzoni. Safe: Self- attentive function embeddings for binary similarity. Detec- tion of Intrusions and Malware, and Vulnerability Assess- ment, pages 309–329, 2019. 5, 6

    2019

  16. [24]

    Self-supervised distilled learning for multi-modal mis- information identification

    Michael Mu, Sreyasee Das Bhattacharjee, and Junsong Yuan. Self-supervised distilled learning for multi-modal mis- information identification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 2819–2828, 2023. 1

    2023

  17. [25]

    The covid-19 ‘info- demic’: A new front for information professionals

    Salman Bin Naeem and Rubina Bhatti. The covid-19 ‘info- demic’: A new front for information professionals. Health Information & Libraries Journal, 37(3):233–239, 2020. 1

    2020

  18. [26]

    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 Neural Information Processing Systems , 35...

    2022

  19. [27]

    Synthetic mis- informers: Generating and combating multimodal misinfor- mation

    Stefanos-Iordanis Papadopoulos, Christos Koutlis, Symeon Papadopoulos, and Panagiotis Petrantonakis. Synthetic mis- informers: Generating and combating multimodal misinfor- mation. In Proceedings of the 2nd ACM International Work- shop on Multimedia AI against Disinformation, ...

  20. [28]

    Red-dot: Multimodal fact-checking via relevant evidence detection

    Stefanos-Iordanis Papadopoulos, Christos Koutlis, Symeon Papadopoulos, and Panagiotis C Petrantonakis. Red-dot: Multimodal fact-checking via relevant evidence detection. arXiv preprint arXiv:2311.09939, 2023. 1, 8

    2023 arXiv

  21. [29]

    Verite: A robust benchmark for multimodal misinformation detection accounting for unimodal bias

    Stefanos-Iordanis Papadopoulos, Christos Koutlis, Symeon Papadopoulos, and Panagiotis C Petrantonakis. Verite: A robust benchmark for multimodal misinformation detection accounting for unimodal bias. International Journal of Mul- timedia Information Retrieval, 13(1):4, 2024. 8

    2024

  22. [30]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, and other. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems , 32,

  23. [31]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13052–13062, 2024. 1, 2, 3, 4, 5, 6, 8

    2024

  24. [32]

    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 International Conference on Machine Learning...

    2021

  25. [33]

    Detecting and grounding multi-modal media manip- ulation and beyond

    Rui Shao, Tianxing Wu, Jianlong Wu, Liqiang Nie, and Zi- wei Liu. Detecting and grounding multi-modal media manip- ulation and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1

    2024

  26. [34]

    Prompt- ing large language models with answer heuristics for knowledge-based visual question answering

    Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu. Prompt- ing large language models with answer heuristics for knowledge-based visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 14974–14983, 2023. 5

    2023

  27. [35]

    Multimodal misinformation detection using large vision- language models

    Sahar Tahmasebi, Eric M ¨uller-Budack, and Ralph Ewerth. Multimodal misinformation detection using large vision- language models. In Proceedings of the 33rd ACM Interna- tional Conference on Information and Knowledge Manage- ment, pages 2189–2199, 2024. 2

    2024

  28. [36]

    Deepface: Closing the gap to human-level perfor- mance in face verification

    Yaniv Taigman, Ming Yang, Marc’Aurelio Ranzato, and Lior Wolf. Deepface: Closing the gap to human-level perfor- mance in face verification. In Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 1701–1708, 2014. 1

    2014

  29. [37]

    Gemini: A family of highly capable multimodal models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean- Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2

    2023 arXiv

  30. [38]

    Mmidr: Teaching large language model to interpret multimodal mis- information via knowledge distillation

    Longzheng Wang, Xiaohan Xu, Lei Zhang, Jiarui Lu, Yongxiu Xu, Hongbo Xu, and Chuang Zhang. Mmidr: Teaching large language model to interpret multimodal mis- information via knowledge distillation. arXiv preprint arXiv:2403.14171, 2024. 1, 3

    2024 arXiv

  31. [39]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 4, 5, 6, 7

    2024 arXiv

  32. [40]

    Eann: Event ad- versarial neural networks for multi-modal fake news detec- tion

    Yaqing Wang, Fenglong Ma, Zhiwei Jin, Ye Yuan, Guangxu Xun, Kishlay Jha, Lu Su, and Jing Gao. Eann: Event ad- versarial neural networks for multi-modal fake news detec- tion. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, p...

    2018

  33. [41]

    Cotkr: Chain-of-thought en- hanced knowledge rewriting for complex knowledge graph question answering

    Yike Wu, Yi Huang, Nan Hu, Yuncheng Hua, Guilin Qi, Jiaoyan Chen, and Jeff Z Pan. Cotkr: Chain-of-thought en- hanced knowledge rewriting for complex knowledge graph question answering. In Proceedings of the 2024 Confer- ence on Empirical Methods in Natural Language Processing ...

    2024

  34. [42]

    Gpt4tools: Teaching large language model to use tools via self-instruction

    Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction. Advances in Neural Information Processing Systems, 36, 2024. 3

    2024

  35. [43]

    A survey on multimodal large language models

    Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 2

    2023 arXiv

  36. [44]

    Support or refute: Analyzing the stance of evidence to detect out-of-context mis-and disinformation

    Xin Yuan, Jie Guo, Weidong Qiu, Zheng Huang, and Shu- jun Li. Support or refute: Analyzing the stance of evidence to detect out-of-context mis-and disinformation. In Proceed- ings of the 2023 Conference on Empirical Methods in Natu- ral Language Processing (EMNLP 2023), pages ...

    2023

  37. [45]

    Ecenet: Explainable and context- enhanced network for muti-modal fact verification

    Fanrui Zhang, Jiawei Liu, Qiang Zhang, Esther Sun, Jingyi Xie, and Zheng-Jun Zha. Ecenet: Explainable and context- enhanced network for muti-modal fact verification. In Pro- ceedings of the 31st ACM International Conference on Mul- timedia, pages 1231–1240, 2023. 1, 3, 5, 6

    2023

  38. [46]

    Aligning instruction tasks unlocks large language models as zero-shot relation extractors

    Kai Zhang, Bernal Jimenez Gutierrez, and Yu Su. Aligning instruction tasks unlocks large language models as zero-shot relation extractors. In Proceedings of the 61st Annual Meet- ing of the Association For Computational Linguistics, 2023. 3

    2023

  39. [47]

    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. 5, 6

    2023 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.