Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

E-FreeM2: Efficient Training-Free Multi-Scale and Cross-Modal News Verification via MLLMs

T0 review · 5 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 training-free retrieval pipeline with two-stage multimodal model reasoning reports 90.0% accuracy on NewsCLIPpings, beating trained baselines.

desk verdict Plausible training-free retrieval pipeline with a headline number that is not yet auditable; the core idea is worth a referee, but the paper overclaims on benchmarks, attacks, and edge deployment. read the letter →

arxiv 2506.20944 v1 pith:JVL3FIUB submitted 2025-06-26 cs.MM cs.CR

Van-Hoang Phan , Long-Khanh Pham , Dang Vu , Anh-Duy Tran , Minh-Son Dao This is my paper
classification cs.MMcs.CR
keywords out-of-contextmisinformationtraining-freeverificationmultimodallargelanguagemodelsretrieval-augmentedfact-checkingchain-of-thoughtpromptingNewsCLIPpingsvisual-centricrankingcross-modalevidenceretrieval
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 claims that out-of-context image-fact checking, which the paper treats as detecting when a genuine photo is paired with a misleading caption, can be done as accurately without any training as with it. On the NewsCLIPpings benchmark, its retrieval-and-reasoning system, E-FreeM2, reports 90.0% overall accuracy, above the trained SNIFFER (88.8%), CCN (84.7%), and DT-Transformer (77.1%) baselines, while using zero trainable parameters and no training data. The system works by retrieving text and visually similar images from the web, filtering them for similarity and source trust, ranking by visual match, and then feeding the shortlist to two off-the-shelf multimodal LLMs in a chain-of-thought sequence that ends in a verdict and an explanation. If accurate, this makes state-of-the-art verification a matter of evidence quality and prompting rather than model training, with a pipeline light enough to run on edge devices.

What carries the argument

The load-bearing mechanism is the two-stage chain-of-thought decision protocol, because it converts retrieved evidence into a verdict without any learned parameters. In Stage 1, a multimodal LLM (Gemini) checks the alignment between the news caption and each candidate's text and metadata, attending to entities, time, place, and consistency; in Stage 2, another model (GPT-4o mini) merges that evidence with its own reading of the pixels to deliver the final out-of-context classification and a confidence score. The retrieval stage feeds this protocol with candidates ranked by visual similarity (S_final = S_visual), on the paper's rationale that a visually similar image is more likely to preserve the true context of the event than a text-similar but unrelated image. The filter module's domain check removes untrustworthy sources, which the ablation shows is essential: low-quality evidence actively degrades the final judgment.

What would settle it

Take a set of fresh out-of-context pairs created after the models' training cutoff, with captions paraphrased so they differ from public benchmark text, and run the same pipeline untouched; if accuracy on those new items falls toward chance while the original NewsCLIPpings score stays at 90%, the benchmark result is driven by memorized or leaked answers rather than by the pipeline's retrieval and reasoning. A simpler check: disable the retrieval stage entirely and measure the drop—the paper reports that removing one evidence type costs about 13 points, but a no-retrieval condition would reveal how much of the 90% depends on live evidence.

Watch

Extended reading notes

Core claim

The central discovery claimed is that a training-free pipeline—live web retrieval plus prompted reasoning by large multimodal language models—reaches 90.0% accuracy on the NewsCLIPpings Merged/Balance test set, outperforming every trained baseline the paper compares against. The design choices are a multi-scale cross-modal retrieval stage (claim-conditioned text search plus image reverse search), a three-part filter (similarity threshold θ=0.7, trusted-domain and English language checks, redundancy removal), and a visual-centric ranking rule that uses visual similarity as the final candidate score. Verification then proceeds in two prompted stages: Gemini first judges whether each retrieved candidate does or does not support the caption's claim, and GPT-4o mini then synthesizes those judgments with direct visual inspection to decide whether the image is out-of-context, producing an explanation and a 0–10 confidence score. The claimed ablations show the domain filter is the largest single component of the gain (without it, accuracy falls to 56.46%), with both visual and textual evidence contributing about 13 points each when removed.

Load-bearing premise

The result depends on the two closed API models (Gemini in Stage 1, GPT-4o mini in Stage 2) genuinely reasoning from the retrieved evidence rather than recalling the benchmark's images or captions from their pretraining, and on the search engine actually returning relevant current pages for every test item.

Editorial extensions

If this is right

  • On the paper's evidence, a completely training-free system can outscore trained detectors on a standard OOC benchmark, so the bottleneck for this task may be evidence retrieval and prompt design rather than learned model capacity.
  • Because the pipeline stores no training data and learns nothing, it can track shifting misinformation by simply re-running retrieval against current web content, avoiding the staleness and poisoning exposure of fixed training datasets.
  • The system outputs a natural-language explanation and a confidence score for every verdict, which the paper presents as an interpretability advantage over black-box trained classifiers.
  • Domain filtering is not a minor detail: the ablation shows it contributes roughly 33.5 points (90.0% down to 56.46% without it), so the quality and trustworthiness of retrieved sources is load-bearing for the method.
  • With zero trainable parameters and an inference time of 12.77 seconds per sample, the paper positions the system as suitable for on-device fact-checking in mobile and wireless settings.

Reading between the lines

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

  • If retrieval quality is what mostly drives the score, a comparable open-weights vision-language model with the same retrieval and prompts should land near the same accuracy; running that experiment would make the result cheaper, auditable, and reproducible without closed APIs.
  • The visual-centric ranking rule (visual similarity as the final score) may be brittle for text-critical manipulations, where the misleading claim mentions a different entity than the visually similar image; a dynamic weighting of visual and textual similarity, conditioned on how strongly they agree, is a natural variant to test.
  • The live-search dependency means the system's accuracy on future, unseen events is bounded by the search engine's coverage and recency; substituting a curated, timestamped news corpus as the retrieval index would let the authors measure how much of the performance comes from the quality of the index rather than the models.
  • The paper's security framing suggests resistance to training-data poisoning, which is structurally true because there is no training set to poison; but the retrieval index and the closed models' priors remain attack surfaces, so a poisoning study on the search results is an open test.
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

5 major / 5 minor

Summary. E-FreeM2 is a training-free, retrieval-based out-of-context misinformation detection system. It retrieves external textual and visual evidence via search engines, filters candidates using modality-specific similarity thresholds and domain heuristics, and finally classifies image-caption pairs through a two-stage Chain-of-Thought procedure using Gemini (Stage 1) and GPT-4o mini (Stage 2). The paper claims 90.0% overall accuracy on the NewsCLIPpings Merged/Balance test set, outperforming trained baselines such as SNIFFER (88.8% in text, 88.4% in Table 1), CCN, and DT-Transformer, while requiring zero trainable parameters. Ablation results show that removing image evidence, text evidence, or domain filters degrades accuracy to 76.48%, 77.05%, and 56.46%, respectively. The paper also promises evaluation on two benchmarks and robustness against attack vectors, but the experimental section covers only one benchmark and no attack robustness experiments.

Significance. If the 90.0% accuracy claim is valid and attributable to the proposed retrieval-and-reasoning pipeline, the result is significant: a completely training-free system beating trained state-of-the-art methods on a standard out-of-context benchmark, with zero trainable parameters and generated explanations, would be an important step for lightweight, adaptable misinformation detection. The paper's core strengths are the clear pipeline design and the explicit focus on training-free efficiency, which contrasts with the expensive training paradigms of prior work. However, the significance is contingent on the auditability of the evaluation: the reliance on closed API models, unspecified prompts, and a public benchmark that may overlap with those models' pretraining data means that the central accuracy claim is not yet verifiable. The ablation study provides some internal consistency but also reveals a strong benchmark prior even without external evidence, which must be contextualized with a no-retrieval control.

major comments (5)
  1. [Abstract, §4.1.1] The abstract and introduction claim 'experiments on two fact-checking benchmarks achieve SOTA results' and 'robustness against various attack vectors,' but §4.1.1 evaluates only the NewsCLIPpings Merged/Balance test set, and no adversarial attack experiments appear in the paper. The claims must either be substantiated by additional experiments or removed/rephrased to match the actual evaluation.
  2. [§4.2.1 and Table 1] The SNIFFER baseline is reported as 88.4 in Table 1 but as 88.8 in the text of §4.2.1. This internal inconsistency is in the central comparison that supports the SOTA claim. Additionally, the paper reports no confidence intervals, significance tests, or multiple trial statistics for the 90.0% accuracy; with 7,264 test samples, the difference from SNIFFER may be within sampling error. Please provide error bars or at least a statistical significance test.
  3. [§3.2] The final decisions are made by proprietary closed models (Gemini and GPT-4o mini), and the paper does not state model versions, snapshot dates, prompts, temperatures, or any retrieval logs. Because NewsCLIPpings is a public benchmark derived from web articles, the 90.0% figure could partly reflect benchmark memorization during pretraining or reverse-image-search leakage that recovers the original source context, rather than the proposed cross-modal reasoning. The authors should release the exact prompts and model versions, and should run a no-retrieval control on the same test items (i.e., the MLLMs alone with only the image-caption pair) to demonstrate that the retrieval pipeline adds value beyond the MLLMs' priors. Table 3 shows that without image or text evidence accuracy is about 76-77%, but this is not the same as a fully no-retrieval control, since the model still sees the other evidence modality.
  4. [§4.3 and Eq. (1)] The similarity threshold θ=0.7, the visual-centric ranking weight, and the top-k candidate count are hyperparameters whose effects are evaluated directly on the same test split (Section 4.3). Since no validation split is used and the method is training-free, selecting these values based on test performance risks overfitting the reported 90.0% to the test set. The paper should describe how these values were chosen and report sensitivity to them across a reasonable range.
  5. [§4.3.3, Table 3] The ablation in Table 3 shows that removing image or text evidence leaves accuracy at 76.48% and 77.05%, respectively, which is still close to or above several trained baselines. This suggests that the underlying MLLMs alone already capture a large portion of the signal, making the specific contribution of the multi-scale retrieval pipeline unclear. The paper should report the accuracy of the two-stage CoT pipeline with no external retrieval at all, and should discuss how much of the 90.0% is attributable to retrieval versus the MLLMs' intrinsic reasoning on the NewsCLIPpings distribution.
minor comments (5)
  1. [§4.2.2] The text says 'our approach achieves an inference time of 12.77 seconds ... as shown in Table 1' but Table 1 contains no inference time and Table 2 lists only trainable parameters. Please add the inference time to the appropriate table or state it only in the text.
  2. [Table 3] The table caption says 'Evaluation results for different evidence filtering methods,' but the rows are ablations of evidence types (image, text, domain filters). Please adjust the caption to accurately describe the content.
  3. [Figure 2] The description of the ranking/filter ablation in §4.3.1 is very brief and does not specify which similarity model, which domain list, or how top-k candidates were selected. This figure is not reproducible from the text.
  4. [§5] The limitation section mentions reliance on external search engines but does not mention the cost or rate limits of commercial search APIs and closed MLLM APIs, which are relevant to the 'lightweight edge device' claim. Please clarify the practical cost and dependency assumptions.
  5. [References] Several references have incomplete bibliographic information (e.g., missing venue or DOIs) and reference [13] is cited as both NewsCLIPpings and 'Newsclippings: Automatic generation...' with inconsistent capitalization. A thorough reference cleanup is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported 90.0% is an external benchmark measurement, not a construction from the method's own definitions; the paper's risks are auditability and test-set selection, not circular derivation.

full rationale

E-FreeM2 does not present a derivation whose conclusion is assumed in its premises. The headline 90.0% accuracy (Section 4.2.1) is a measured outcome on the fixed NewsCLIPpings Merged/Balance test set, not a quantity constructed from the method's definitions: the retrieval, the filtering threshold (theta = 0.7, Eq. 1), the visual-centric ranking (Eq. 2), and the two-stage prompting (Section 3.2) are all defined independently of the test labels. The ablations in Section 4.3 compare configurations on the same split, which can inflate optimism, but the paper labels the threshold as 'predefined' and reports no fitting procedure; test-set reuse is a validity/selection concern, not a circular equivalence. Self-citations [18] and [30] appear only in related-work enumerations and are not load-bearing for the SOTA claim; no uniqueness theorem or ansatz is imported from prior work. The main weaknesses are auditability (closed API models, prompts/retrieval logs not released) and the abstract's claim of two benchmarks while only NewsCLIPpings is evaluated in Section 4.1.1; these are reproducibility and completeness problems rather than circularity. No step in the paper reduces by construction to its own input.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

Total ledger: three free parameters (threshold, visual-only ranking, unreported top-k count) plus five domain assumptions. The system's reported advantage depends on these choices: theta=0.7 and the trusted-domain whitelist are untested inputs, and the final accuracy rests on unverified assumptions about search quality, embedding validity, MLLM reliability, and benchmark non-contamination. No invented entities are introduced.

free parameters (3)
  • Similarity threshold theta = 0.7
    Equation (1) keeps candidates with S_similarity >= theta, but theta=0.7 is asserted without derivation or validation on a separate split; ablation in Section 4.3 suggests components were tuned on the same test benchmark.
  • Visual ranking weight = S_final = S_visual (visual-only)
    Equation (2) sets the final retrieval score entirely to visual similarity, a hand-chosen heuristic acknowledged in Section 5 as potentially suboptimal when textual cues dominate.
  • Top-k candidate count = not reported
    Ablation varies Top-1, Top-2, and Top-3, but the actual number of evidence candidates fed to the LLM in the final system is not specified, leaving a key pipeline parameter uncontrolled.
assumptions (5)
  • domain assumption External search engines return relevant and temporally appropriate evidence for each query.
    Section 3.1 builds both retrieval pipelines on search results, and Section 5 concedes dependence on availability and quality of external sources.
  • domain assumption The curated list of trusted news domains is correct and complete.
    Domain Filtering in Section 3.1 assumes a whitelist of reputable sources; ablation (Section 4.3.3) shows removing it drops accuracy to 56.46%, but the list is not released.
  • domain assumption Dense-embedding cosine similarities (all-MiniLM-L6-v2 for text, ViT for images) accurately measure task-relevant relevance.
    Section 3.1 uses these scores as the sole relevance metric and does not calibrate them against human judgments.
  • domain assumption Gemini and GPT-4o mini provide reliable fact-verification judgments under the given prompts.
    Section 3.2 delegates all verification and final decisions to these closed models with no calibration or error analysis.
  • domain assumption NewsCLIPpings test labels are ground truth and have not been memorized by the proprietary MLLMs.
    The SOTA claim in Section 4.2 presumes a clean held-out evaluation, but the public benchmark likely appeared in pretraining data; this is never discussed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of E-FreeM2: Efficient Training-Free Multi-Scale and Cross-Modal News Verification via MLLMs." pith.science (2026). https://pith.science/paper/JVL3FIUB

@misc{pith2026250620944,
  author       = {Pith},
  title        = {Pith review of: E-FreeM2: Efficient Training-Free Multi-Scale and Cross-Modal News Verification via MLLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JVL3FIUB}},
  note         = {Machine review of arXiv:2506.20944}
}
read the original abstract

The rapid spread of misinformation in mobile and wireless networks presents critical security challenges. This study introduces a training-free, retrieval-based multimodal fact verification system that leverages pretrained vision-language models and large language models for credibility assessment. By dynamically retrieving and cross-referencing trusted data sources, our approach mitigates vulnerabilities of traditional training-based models, such as adversarial attacks and data poisoning. Additionally, its lightweight design enables seamless edge device integration without extensive on-device processing. Experiments on two fact-checking benchmarks achieve SOTA results, confirming its effectiveness in misinformation detection and its robustness against various attack vectors, highlighting its potential to enhance security in mobile and wireless communication environments.

Figures

Figures reproduced from arXiv: 2506.20944 by the authors.

Figure 1
Figure 1. Overview of E-FreeM2 Framework. Our efficient training-free multi-scale and cross-modal news verification system operates through two main components: (1) Multi-Scale Cross-Modal Evidence Retrieval - employing dual textual and visual retrieval pipelines with hierarchical similarity-based filtering and visual-centric ranking to obtain contextually relevant evidence from external sources; and (2) Two-Stage Multimodal … view at source ↗
Figure 2
Figure 2. Ablation results with filter strategies 4.3.2 Training Dataset Utilization. To compare the effectiveness of our approach against SNIFFER, [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Percentage of used training set of NewsCLIPpings[13] Method Accuracy E-FreeM2 (Ours) 90.0 - w/o Image Evidences 76.48 - w/o Text Evidences 77.05 - w/o Domain Filters 56.46 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Multimedia Verification Through Multi-Agent Deep Research Multimodal Large Language Models

    cs.CV 2025-07 conditional novelty 4.0 of 10

    A six-stage multi-agent MLLM pipeline with reverse image search, metadata analysis, and fact-checking tools is demonstrated on a single Ukraine missile-strike video, with no quantitative evaluation.

Reference graph

Works this paper leans on

43 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    Sahar Abdelnabi, Rakibul Hasan, and Mario Fritz. 2022. 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 . 14940–14949

    work page 2022

  2. [2]

    Nadia Alonso-López, Pavel Sidorenko Bautista, and Fábio Giacomelli. 2021. Be- yond Challenges and Viral Dance Moves: TikTok as a Vehicle for Disinformation and Fact-Checking in Spain, Portugal, Brazil, and the USA. Anàlisi 64 (06 2021), 65–84. https://doi.org/10.5565/rev/analisi.3411>

    work page doi:10.5565/rev/analisi.3411 2021

  3. [3]

    Shivangi Aneja, Chris Bregler, and Matthias Nießner. 2023. COSMOS: catching out-of-context image misuse using self-supervised learning. In Proceedings of the AAAI conference on artificial intelligence , Vol. 37. 14084–14092

    work page 2023

  4. [4]

    Sarah Evanega, Mark Lynas, Jordan Adams, and Karinne Smolenyak. 2020. Coro- navirus misinformation: quantifying sources and themes in the COVID-19 ‘info- demic’ (Preprint). https://doi.org/10.2196/preprints.25143

    work page doi:10.2196/preprints.25143 2020

  5. [5]

    Lisa Fazio. 2020. Out-of-context photos are a powerful low-tech form of misin- formation. The Conversation 14, 1 (2020)

    work page 2020

  6. [6]

    Yimeng Gu, Mengqi Zhang, Ignacio Castro, Shu Wu, and Gareth Tyson. 2024. Learning Domain-Invariant Features for Out-of-Context News Detection. arXiv preprint arXiv:2406.07430 (2024)

    work page Pith review arXiv 2024

  7. [7]

    Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, et al. 2023. Lan- guage is not all you need: Aligning perception with language models. Advances in Neural Information Processing Systems 36 (2023), 72096–72109

    2023

  8. [8]

    Ayush Jaiswal, Ekraam Sabir, Wael AbdAlmageed, and Premkumar Natarajan

    work page

Show all 43 references
  1. [9]

    Ayush Jaiswal, Yue Wu, Wael AbdAlmageed, Iacopo Masi, and Premkumar Natara- jan. 2019. Aird: Adversarial learning framework for image repurposing detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. 11330–11339

    2019

  2. [10]

    Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang

  3. [11]

    Yihui Li, Xiaoyue Yan, Hao Zhou, and Borong Lin. 2024. Question Answering for Decisionmaking in Green Building Design: A Multimodal Data Reasoning Method Driven by Large Language Models. arXiv preprint arXiv:2412.04741 (2024)

    2024

  4. [12]

    Piper Liu and Vincent Huang. 2020. Digital Disinformation About COVID-19 and the Third-Person Effect: Examining the Channel Differences and Negative Emotional Outcomes. Cyberpsychology, Behavior, and Social Networking 23 (07 2020). https://doi.org/10.1089/cyber.2020.0363

    2020

  5. [13]

    Grace Luo, Trevor Darrell, and Anna Rohrbach. 2021. Newsclippings: Automatic generation of out-of-context multimodal media. arXiv preprint arXiv:2104.05893 (2021)

    2021 arXiv

  6. [14]

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

    2023

  7. [15]

    Alex Nikolov, Giovanni Da San Martino, Ivan Koychev, and Preslav Nakov. 2020. Team Alex at CLEF CheckThat! 2020: Identifying Check-Worthy Tweets With Transformer Models. arXiv:2009.02931 [cs.CL] https://arxiv.org/abs/2009.02931

    2020 arXiv

  8. [16]

    Stefanos-Iordanis Papadopoulos, Christos Koutlis, Symeon Papadopoulos, and Panagiotis Petrantonakis. 2023. Synthetic misinformers: Generating and com- bating multimodal misinformation. In Proceedings of the 2nd ACM International Workshop on Multimedia AI against Disinformation . 36–44

    2023

  9. [17]

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

    2023 arXiv

  10. [18]

    Long-Khanh Pham, Hoa-Vien Vo-Hoang, and Anh-Duy Tran. 2024. A Generative Adaptive Context Learning Framework for Large Language Models in Cheapfake Detection (ICMR ’24). Association for Computing Machinery, New York, NY, USA, 1288–1293. https://doi.org/10.1145/3652583.3657597

    2024

  11. [19]

    Gwenyth Portillo Wightman, Alexandra Delucia, and Mark Dredze. 2023. Strength in Numbers: Estimating Confidence of Large Language Models by Prompt Agreement. In Proceedings of the 3rd Workshop on Trustworthy Natural Language Processing (TrustNLP 2023), Anaelia Ovalle, Kai-Wei ...

    2023

  12. [20]

    Shraman Pramanick, Rama Chellappa, and Subhashini Venugopalan. 2025. SPIQA: A Dataset for Multimodal Question Answering on Scientific Papers. arXiv:2407.09413 [cs.CL] https://arxiv.org/abs/2407.09413

    2025 arXiv

  13. [21]

    Peng Qi, Zehong Yan, Wynne Hsu, and Mong Li Lee. 2024. 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 . 13052–13062

    2024

  14. [22]

    Shuofei Qiao, Yixin Ou, Ningyu Zhang, Xiang Chen, Yunzhi Yao, Shumin Deng, Chuanqi Tan, Fei Huang, and Huajun Chen. 2023. Reasoning with Language Model Prompting: A Survey. arXiv:2212.09597 [cs.CL] https://arxiv.org/abs/ 2212.09597

    2023 arXiv

  15. [23]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  16. [24]

    Ekraam Sabir, Wael AbdAlmageed, Yue Wu, and Prem Natarajan. 2018. Deep multimodal image-repurposing detection. In Proceedings of the 26th ACM inter- national conference on Multimedia . 1337–1345

    2018

  17. [25]

    Ian Sample. 2020. What are deepfakes – and how can you spot them? Re- trieved Mar 08, 2024 from https://www.reuters.com/world/middle-east/false- claims-israel-hamas-war-mushroom-online-put-focus-musks-x-2023-10-10/

    2020

  18. [26]

    Rui Shao, Tianxing Wu, and Ziwei Liu. 2023. Detecting and grounding multi- modal media manipulation. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition . 6904–6913

    2023

  19. [27]

    Sahar Tahmasebi, Eric Müller-Budack, and Ralph Ewerth. 2024. Multimodal misinformation detection using large vision-language models. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2189–2199

    2024

  20. [28]

    Ruben Tolosana, Ruben Vera-Rodriguez, Julian Fierrez, Aythami Morales, and Javier Ortega-Garcia. 2020. Deepfakes and beyond: A survey of face manipulation and fake detection. Information Fusion 64 (2020), 131–148

    2020

  21. [29]

    Image, Tell me your story!

    Jonathan Tonglet, Marie-Francine Moens, and Iryna Gurevych. 2024. “Image, Tell me your story!” Predicting the original meta-context of visual misinfor- mation. In Proceedings of the 2024 Conference on Empirical Methods in Natu- ral Language Processing, Yaser Al-Onaizan, Mohit ...

    2024

  22. [30]

    Hoa-Vien Vo-Hoang, Long-Khanh Pham, and Minh-Son Dao. 2024. Detecting Out- of-Context Media with LLaMa-Adapter V2 and RoBERTa: An Effective Method for Cheapfakes Detection. In Proceedings of the 2024 International Conference on Multimedia Retrieval (Phuket, Thailand) (ICMR ’24...

    2024 doi

  23. [31]

    Xueyu Wang, Jiajun Huang, Siqi Ma, Surya Nepal, and Chang Xu. 2022. Deepfake disrupter: The detector of deepfake is my friend. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 14920–14929

    2022

  24. [32]

    Yaqing Wang, Fenglong Ma, Zhiwei Jin, Ye Yuan, Guangxu Xun, Kishlay Jha, Lu Su, and Jing Gao. 2018. Eann: Event adversarial neural networks for multi-modal fake news detection. In Proceedings of the 24th acm sigkdd international conference on knowledge discovery & data mining ...

    2018

  25. [33]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903 [cs.CL] https: //arxiv.org/abs/2201.11903

    2023 arXiv

  26. [34]

    Tianhe Wu, Kede Ma, Jie Liang, Yujiu Yang, and Lei Zhang. 2024. A Comprehen- sive Study of Multimodal Large Language Models for Image Quality Assessment. arXiv:2403.10854 [cs.CV] https://arxiv.org/abs/2403.10854

    2024 arXiv

  27. [35]

    Haotian Xia, Zhengbang Yang, Junbo Zou, Rhys Tracy, Yuqing Wang, Chi Lu, Christopher Lai, Yanjun He, Xun Shao, Zhuoqing Xie, Yuan fang Wang, Weining Shen, and Hanjie Chen. 2024. SPORTU: A Comprehensive Sports Understanding Benchmark for Multimodal Large Language Models. arXiv:...

    2024 arXiv

  28. [36]

    Xin Yuan, Jie Guo, Weidong Qiu, Zheng Huang, and Shujun Li. 2023. Support or refute: Analyzing the stance of evidence to detect out-of-context mis-and disinformation. arXiv preprint arXiv:2311.01766 (2023)

    2023 arXiv

  29. [37]

    Fanrui Zhang, Jiawei Liu, Qiang Zhang, Esther Sun, Jingyi Xie, and Zheng-Jun Zha. 2023. Ecenet: explainable and context-enhanced network for muti-modal fact verification. In Proceedings of the 31st ACM International Conference on Multimedia. 1231–1240

    2023

  30. [38]

    Yizhou Zhang, Loc Trinh, Defu Cao, Zijun Cui, and Yan Liu. 2024. Interpretable Detection of Out-of-Context Misinformation with Neural-Symbolic-Enhanced Large Multimodal Model. arXiv:2304.07633 [cs.CL] https://arxiv.org/abs/2304. 07633

    2024 arXiv

  31. [39]

    Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. 2021. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 2185–2194. SCID ’25, August 25–29, 2025, Hanoi, Vietnam Van-Ho...

    2021

  32. [40]

    Xinyi Zhou, Jindi Wu, and Reza Zafarani. 2020. : Similarity-aware multi-modal fake news detection. In Pacific-Asia Conference on knowledge discovery and data mining. Springer, 354–367

    2020

  33. [2017]

    In Proceedings of the 25th ACM international conference on Multimedia

    Multimedia semantic integrity assessment using joint embedding of images and text. In Proceedings of the 25th ACM international conference on Multimedia . 1465–1471

  34. [2019]

    arXiv preprint arXiv:1908.03557 (2019)

    Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557 (2019)

    2019 arXiv

  35. [7864]

    https://doi.org/10.18653/v1/2024.emnlp-main.448

    2024 doi

Pith tools

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