Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Mitigating GenAI-powered Evidence Pollution for Out-of-Context Multimodal Misinformation Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read AI-generated evidence trips up misinformation detectors by 9 points

desk verdict A timely first look at multimodal evidence pollution, but the headline numbers are confounded by evidence volume and truncation; the paper needs two control baselines before its causal claims hold. read the letter →

arxiv 2501.14728 v1 pith:ESUC37GQ submitted 2025-01-24 cs.MM cs.CLcs.CVcs.CY

Zehong Yan , Peng Qi , Wynne Hsu , Mong Li Lee This is my paper · ORCID
classification cs.MMcs.CLcs.CVcs.CY
keywords out-of-contextmisinformationmultimodalfact-checkingevidencepollutionGenAI-generatedtextcross-modalrerankingclaim-evidencereasoningCLIPembeddingsrobustness
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

Existing out-of-context (OOC) misinformation detectors assume that the web evidence they retrieve to check a news image is factual; the paper shows that mixing in GenAI-generated text and images drops their accuracy by more than 9 percentage points. The authors build a pollution pipeline with GPT-4 zero-shot text generation and depth-conditioned Stable Diffusion image generation, and demonstrate the drop across four detectors: CCN, RED-DOT, SNIFFER, and GPT-4o. They then propose two plug-and-play strategies—cross-modal evidence reranking and cross-modal claim-evidence reasoning—that restore accuracy close to the clean baseline without retraining. A sympathetic reader would take this as evidence that evidence-level pollution, not just stylistic claim rewriting, is a real and fixable vulnerability for multimodal fact-checking.

What carries the argument

The load-bearing object is CLIP's shared image-text embedding space, used in two passes. Cross-modal evidence reranking scores each retrieved textual evidence by cosine similarity to the claim image, and each retrieved visual evidence by similarity to the claim caption, then keeps only the top-k ranked items. Cross-modal claim-evidence reasoning finds the retrieved text most similar to the caption, and passes that text together with the claim image into the detector's existing consistency-checking module for a second opinion; this catches false captions that describe the image correctly but attach it to the wrong event. Because both stages act on the detector's input rather than its weights, they plug into CCN, RED-DOT, SNIFFER, and GPT-4o without retraining.

What would settle it

Run the same four detectors on evidence retrieved from the live web at a time when GenAI-generated content is naturally present, comparing accuracy against a clean historical retrieval; if the accuracy drop is not reproduced, or the two strategies do not recover it, the vulnerability and the defense are artifacts of the synthetic pollution pipeline.

Watch

Extended reading notes

Core claim

The paper's central claim is that GenAI-generated evidence injected into the retrieval corpus is a serious, currently unhandled threat to OOC detection, and that reranking evidence by cross-modal semantic relevance plus a secondary image-to-text consistency check largely neutralizes it. Using GPT-4 to generate entity-based, supporting, and refuting textual evidence and Depth-Conditional Stable Diffusion to create visually varied claim-conditioned images, the authors show that combined text-and-image pollution lowers accuracy by 9.55–12.50 points on NewsCLIPpings and 11.33–15.54 points on VERITE across all four tested detectors. With both proposed strategies active, SNIFFER recovers from 76.42% to 88.82% on NewsCLIPpings and from 59.41% to 72.82% on VERITE, close to its clean baselines of 88.85% and 73.69%. The human and model evaluations also show that generated evidence is hard to identify—Vicuna-13B detects it at 41.3% accuracy and human participants at 49.39%—so simple filtering is not enough.

Load-bearing premise

The synthetic pollution pipeline of GPT-4 text and depth-conditioned Stable Diffusion images produces evidence that is representative of the GenAI pollution real systems will face, so if actual pollution is more semantically misleading or comes from different models, the measured degradation and recovery may not transfer.

Editorial extensions

If this is right

  • Deployed OOC detectors that rely on web evidence will misclassify a meaningful fraction of claims once GenAI-polluted text and images enter the retrieved corpus, unless they adopt some form of evidence-quality control.
  • Textual pollution causes larger accuracy losses than visual pollution, so systems should weight or cross-check textual evidence rather than trusting it wholesale.
  • The two proposed strategies transfer across detector families without retraining, making them a low-cost upgrade for existing fact-checking pipelines.
  • As image generators improve, the degradation from visual pollution grows, so defenses that only handle today's image quality will age poorly.

Reading between the lines

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

  • If real-world pollution is optimized adversarially to maximize CLIP similarity to the claim, the reranking stage could be neutralized; a worst-case attack would craft evidence that the reranker ranks first, so the strategy is best understood as a robustness improvement, not a guarantee.
  • The defense inherits CLIP's biases, so on specialized domains like medical or financial claims, where CLIP's embeddings are less calibrated, the observed recovery may shrink; this is testable with domain-specific benchmarks.
  • The human-evaluation finding—that reading polluted evidence reverses many correct veracity judgments—suggests the threat extends beyond automated detectors to human fact-checkers, which the paper does not develop.
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 investigates how GenAI-generated evidence pollution affects out-of-context (OOC) multimodal misinformation detectors, which typically retrieve web evidence to assess whether an image is used in a false context. The authors simulate pollution by generating textual evidence with GPT-4 and visual evidence with depth-conditioned Stable Diffusion, and then add this generated evidence to the original clean evidence corpora. Experiments on NewsCLIPpings and VERITE across four detectors (CCN, RED-DOT, SNIFFER, GPT-4o) show accuracy drops of more than 9 percentage points under combined text and image pollution. To mitigate this, the paper proposes two plug-and-play strategies: cross-modal evidence reranking, which uses CLIP to rerank retrieved evidence by similarity to the claim, and cross-modal claim-evidence reasoning, which selects the most caption-relevant textual evidence for a consistency check with the claim image. The reported results indicate that these strategies restore performance close to the clean baseline, with SNIFFER improving from 76.42% to 88.82% accuracy on NewsCLIPpings and from 59.41% to 72.82% on VERITE.

Significance. The problem is timely: as generative models become widespread, the assumption that evidence retrieved for fact-checking is clean is increasingly untenable. The paper constructs a large multimodal pollution corpus, which is a potentially valuable resource, and the proposed strategies are model-agnostic and do not require retraining, giving them practical appeal. The evaluation spans four detectors and two benchmarks, including the real-world VERITE dataset, which strengthens the scope if the results are valid. However, the central quantitative claims currently rest on comparisons that are confounded by evidence volume and by the top-k truncation used in the defense, so the contribution will be fully realized only after the control experiments requested in the major comments are added. The paper also lacks error bars and significance tests, which is an important omission for comparative claims.

major comments (4)
  1. [§3.2, Tables 1–2] The polluted condition adds generated evidence to the clean corpus, roughly doubling the evidence volume (e.g., NewsCLIPpings test text: 60,848 clean vs. 67,016 generated texts; images: 66,772 clean vs. 67,092 generated images). Without a control condition that adds an equal number of factual clean evidence items, the observed accuracy drop cannot be attributed specifically to the false or stance-bearing nature of GenAI output; it could be an effect of corpus size or of adding near-duplicate content. Please add a control that enriches the clean corpus with additional factual evidence to match the polluted corpus size and report detector performance under that condition.
  2. [§4.3, Algorithm 1, Table 3] The defense selects only the top-1 text and top-5 visual evidence after CLIP-based reranking. There is no baseline that applies the same truncation to the polluted corpus in the original retrieval order (or in random order) without CLIP reranking. The reported gains of up to 12.40 accuracy points could therefore be a truncation effect rather than a reranking effect. Please report no-rerank truncation baselines with identical k values, as well as a random-order truncation baseline, to isolate the contribution of the CLIP-based ranking.
  3. [§4.2, Human Evaluation] The human evaluation is based on ten randomly selected misinformation samples. This sample size is far too small to support quantitative statements such as "only 49.39% of the generated evidence was correctly identified as AI-generated" and "41.84% of the initially correct veracity judgments were reversed." Please either substantially increase the sample size or present these numbers as illustrative only, with appropriate caveats and uncertainty.
  4. [Tables 2–4] No error bars, confidence intervals, or statistical significance tests are reported for any of the key comparisons. Because some differences are small and could be within noise (for example, CCN's F1-True on VERITE in Table 3 decreases with Cross-modal Reasoning), the comparative claims are not yet statistically grounded. Please report significance tests, such as bootstrap over test samples or repeated runs with different seeds, for the main comparisons in Tables 2–4.
minor comments (5)
  1. [§4.2] The sentence "Detection of false claims in the presence of with polluted evidence proves to be more challenging" contains a typo ("of with"); please revise.
  2. [Appendix B] There are two typos: "We generate textual polluion" should be "pollution", and "We leverage the the CLIP" has a duplicated article "the".
  3. [§4.3, Table 3] The text states that cross-modal claim-evidence reasoning "shows substantial gains, particularly in the detection of true claims," but Table 3 shows a decrease for CCN on VERITE (F1-True from 68.65 to 63.39 with reasoning alone). Please qualify this claim to acknowledge per-model variation.
  4. [Tables 3–4] The notation "↑12.40" and "↓5.26" is used for percentage-point changes, but the superscript percentage signs may be misread as relative percentages. Please state explicitly that these are absolute percentage-point differences relative to the None condition.
  5. [Appendix F, Table 5] Table 5 reports the fraction of clean evidence within the top-k after CLIP reranking, but it does not show detector accuracy when using those top-k items, so it does not directly demonstrate that reranking improves end-task performance over simple truncation; the connection to the major concern in §4.3 should be made explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical evaluation against external benchmarks, and no derivation reduces to its own inputs.

full rationale

This manuscript is an empirical evaluation against two external benchmarks (NewsCLIPpings and VERITE) and does not derive a target result from its own assumptions. The pollution pipeline (GPT-4 text generation and Depth-Conditional Stable Diffusion image generation) is a data-construction step, not a fitted model, and the reported degradation is measured by running public pretrained detectors (CCN, RED-DOT, SNIFFER, GPT-4o) on the augmented evidence. The proposed defenses are evaluated on held-out test splits without fitting parameters to those splits: cross-modal reranking uses frozen CLIP, and cross-modal claim-evidence reasoning reuses existing consistency modules. The authors do cite their own SNIFFER paper (Qi et al., 2024) as a baseline, but this citation is not load-bearing: SNIFFER is a public checkpoint that is evaluated, not invoked as a premise for the new claims. The only CLIP-related loop is that Appendix C uses CLIP similarity to show generated evidence resembles clean evidence while the reranker also uses CLIP; this is a shared metric, not a reduction by construction, and the paper supports its claims with retrieval-precision numbers (Table 5) and end-task accuracy, which are external measurements. The reviewer concern about missing controls such as equal evidence volume or no-rerank truncation is a question of experimental internal validity and correctness, not circularity. Therefore no circular step meeting the quoted-evidence bar was found.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities or fitted parameters in the sense of a derivation. The free parameters are generation settings and selection thresholds. The axioms are domain assumptions about realism and representativeness, which are typical for empirical ML papers but should be stated explicitly.

free parameters (2)
  • GPT-4 sampling hyperparameters = temperature=1.2, max_tokens=64, top_p=0.95
    These hyperparameters control the generated textual pollution and are chosen by the authors, not justified by a validation procedure.
  • Reranking selection thresholds = top-1 textual evidence, top-5 visual evidence
    The number of evidence items kept after reranking is a design choice that affects performance; no sensitivity analysis or validation-based selection is reported.
assumptions (4)
  • domain assumption The synthetic pollution generated by GPT-4 and Depth-Conditional Stable Diffusion is representative of real-world GenAI evidence pollution.
    The entire evaluation rests on the realism of the simulated pollution. If real pollution differs, the reported degradation and defense effectiveness may not transfer. Invoked in Section 3.2.
  • domain assumption CLIP-based cross-modal similarity is a reliable signal for evidence relevance in this setting.
    The reranking strategy assumes that ranking evidence by CLIP similarity to the other modality selects clean over polluted evidence. The paper provides retrieval-precision numbers but no comparison to other relevance measures. Invoked in Algorithm 1 and Section 3.3.
  • domain assumption The four chosen baselines (CCN, RED-DOT, SNIFFER, GPT-4o) are representative of existing OOC detectors.
    The paper generalizes its conclusions to 'existing OOC detectors' from a small set of models, which is a reasonable but unproven assumption. Invoked in Section 4.1.
  • domain assumption NewsCLIPpings and VERITE benchmarks accurately reflect the OOC misinformation detection task in the wild.
    The evaluation is limited to two benchmarks, one synthetic and one real, both of which may not capture the full diversity of real-world misinformation. Invoked in Section 4.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating GenAI-powered Evidence Pollution for Out-of-Context Multimodal Misinformation Detection." pith.science (2026). https://pith.science/paper/ESUC37GQ

@misc{pith2026250114728,
  author       = {Pith},
  title        = {Pith review of: Mitigating GenAI-powered Evidence Pollution for Out-of-Context Multimodal Misinformation Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ESUC37GQ}},
  note         = {Machine review of arXiv:2501.14728}
}
read the original abstract

While large generative artificial intelligence (GenAI) models have achieved significant success, they also raise growing concerns about online information security due to their potential misuse for generating deceptive content. Out-of-context (OOC) multimodal misinformation detection, which often retrieves Web evidence to identify the repurposing of images in false contexts, faces the issue of reasoning over GenAI-polluted evidence to derive accurate predictions. Existing works simulate GenAI-powered pollution at the claim level with stylistic rewriting to conceal linguistic cues, and ignore evidence-level pollution for such information-seeking applications. In this work, we investigate how polluted evidence affects the performance of existing OOC detectors, revealing a performance degradation of more than 9 percentage points. We propose two strategies, cross-modal evidence reranking and cross-modal claim-evidence reasoning, to address the challenges posed by polluted evidence. Extensive experiments on two benchmark datasets show that these strategies can effectively enhance the robustness of existing out-of-context detectors amidst polluted evidence.

Figures

Figures reproduced from arXiv: 2501.14728 by the authors.

Figure 1
Figure 1. Example of how misinformation detectors are misled by [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An illustrated example of claim-conditioned [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of a typical OOC detection framework. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: OOC misinformation detection framework in the presence of polluted evidence with proposed cross-modal reranking and cross [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: SNIFFER’s performance across varying proportion of pol￾luted evidence and GenAI models on NewsCLIPpings. modal pollution. 2) Textual pollution has a greater impact than visual pollution, indicating that existing OOC detectors are more dependent on textual information. …
Figure 6
Figure 6. Figure 6: Case study of SNIFFER’s justification outputs under clean and polluted settings. The evidence used in the last row is selected [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: An overview of out-of-context detection system under evidence pollution. A claim image and its caption are processed by retrievers [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: (a): Distribution of differences in CLIP scores between input image and textual evidence. The X-axis represents the difference [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: CCN’s performance across varying proportion of polluted [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 11
Figure 11. Figure 11: Prompt used to ask GPT-4o to detect out-of-context misinformation. [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 28 canonical work pages

  1. [1]

    Fact-Saboteurs: A taxonomy of evidence manipula- tion attacks against fact-verification systems

    [Abdelnabi and Fritz, 2023] Sahar Abdelnabi and Mario Fritz. Fact-Saboteurs: A taxonomy of evidence manipula- tion attacks against fact-verification systems. In Joseph A. Calandrino and Carmela Troncoso, editors, 32nd USENIX Security Symposium, USENIX Security 2023, Anaheim, CA, USA, August 9-11, 2023, pages 6719–6736,

    work page 2023

  2. [7]

    FACTIFY3M: A benchmark for multimodal fact verification with explain- ability through 5W question-answering

    [Chakraborty et al., 2023] Megha Chakraborty, Khushbu Pahwa, Anku Rani, Shreyas Chatterjee, Dwip Dalal, Harshit Dave, Ritvik G, Preethi Gurumurthy, Adarsh Ma- hor, Samahriti Mukherjee, Aditya Pakala, Ishan Paul, Janvita Reddy, Arghya Sarkar, Kinjal Sensharma, Aman Chadha, Amit Sheth, and Amitava Das. FACTIFY3M: A benchmark for multimodal fact verification...

    work page 2023

  3. [8]

    Can llm- generated misinformation be detected? In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024,

    [Chen and Shu, 2024] Canyu Chen and Kai Shu. Can llm- generated misinformation be detected? In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024,

    work page 2024

  4. [10]

    Synthetic disinformation attacks on au- tomated fact verification systems

    [Du et al., 2022] Yibing Du, Antoine Bosselut, and Christo- pher D Manning. Synthetic disinformation attacks on au- tomated fact verification systems. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 36, pages 10581–10589,

    work page 2022

  5. [11]

    A survey on automated fact-checking

    [Guo et al., 2022] Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. A survey on automated fact-checking. Transactions of the Association for Computational Lin- guistics, 10:178–206,

    work page 2022

  6. [12]

    Multimedia se- mantic integrity assessment using joint embedding of im- ages and text

    [Jaiswal et al., 2017] Ayush Jaiswal, Ekraam Sabir, Wael AbdAlmageed, and Premkumar Natarajan. Multimedia se- mantic integrity assessment using joint embedding of im- ages and text. In Proceedings of the 25th ACM interna- tional conference on Multimedia, pages 1465–1471,

    work page 2017

  7. [14]

    Aligning large language models through synthetic feedback

    [Kim et al., 2023] Sungdong Kim, Sanghwan Bae, Jamin Shin, Soyoung Kang, Donghyun Kwak, Kang Yoo, and Minjoon Seo. Aligning large language models through synthetic feedback. In Proceedings of the 2023 Confer- ence on Empirical Methods in Natural Language Process- ing, pages 13677–13700, December

    work page 2023

  8. [15]

    VisualBERT: A simple and performant baseline for vision and language

    [Li et al., 2019] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A simple and performant baseline for vision and language. CoRR, abs/1908.03557,

    arXiv 2019

Show all 40 references
  1. [16]

    NewsCLIPpings: Automatic Generation of Out-of-Context Multimodal Media

    [Luo et al., 2021] Grace Luo, Trevor Darrell, and Anna Rohrbach. NewsCLIPpings: Automatic Generation of Out-of-Context Multimodal Media. In Proceedings of the 2021 Conference on Empirical Methods in Natural Lan- guage Processing, pages 6801–6817, November

    2021

  2. [17]

    Multimodal analytics for real-world news using measures of cross-modal entity consistency

    [M¨uller-Budack et al., 2020] Eric M ¨uller-Budack, Jonas Theiner, Sebastian Diering, Maximilian Idahl, and Ralph Ewerth. Multimodal analytics for real-world news using measures of cross-modal entity consistency. In Proceed- ings of the 2020 international conference on multime...

    2020

  3. [18]

    GPT-4 technical report

    [OpenAI, 2023] OpenAI. GPT-4 technical report. CoRR, abs/2303.08774,

    2023 arXiv

  4. [19]

    Hello GPT-4o,

    [OpenAI, 2024] OpenAI. Hello GPT-4o,

    2024

  5. [20]

    [Pan et al., 2023a] Liangming Pan, Wenhu Chen, Min-Yen Kan, and William Yang Wang

    Accessed: 2024-06-07. [Pan et al., 2023a] Liangming Pan, Wenhu Chen, Min-Yen Kan, and William Yang Wang. Attacking open-domain question answering by injecting misinformation. In Pro- ceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd...

    2024

  6. [21]

    On the risk of misinformation pollution with large lan- guage models

    [Pan et al., 2023b] Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Wang. On the risk of misinformation pollution with large lan- guage models. In Findings of the Association for Com- putational Linguistics: EMNLP 2023 , pages 1389–1403, December

    2023

  7. [22]

    Petrantonakis

    [Papadopoulos et al., 2023b] Stefanos-Iordanis Papadopou- los, Christos Koutlis, Symeon Papadopoulos, and Panagio- tis C. Petrantonakis. RED-DOT: multimodal fact-checking via relevant evidence detection. CoRR, abs/2311.09939,

    arXiv

  8. [23]

    VERITE: a robust benchmark for multimodal misinformation detection accounting for unimodal bias

    [Papadopoulos et al., 2024] Stefanos-Iordanis Papadopou- los, Christos Koutlis, Symeon Papadopoulos, and Panagiotis C Petrantonakis. VERITE: a robust benchmark for multimodal misinformation detection accounting for unimodal bias. International Journal of Multimedia Information...

    2024

  9. [24]

    SNIFFER: Multimodal large language model for explainable out-of-context misinformation de- tection

    [Qi et al., 2024] Peng Qi, Zehong Yan, Wynne Hsu, and Mong Li Lee. SNIFFER: Multimodal large language model for explainable out-of-context misinformation de- tection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 13052– 13062,

    2024

  10. [25]

    Learning transferable visual models from nat- ural language supervision

    [Radford et al., 2021] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from nat- ural language supervision. In International confere...

    2021

  11. [26]

    Hierarchi- cal text-conditional image generation with CLIP latents

    [Ramesh et al., 2022] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchi- cal text-conditional image generation with CLIP latents. CoRR, abs/2204.06125,

    2022 arXiv

  12. [27]

    High-resolution image synthesis with latent diffusion models

    [Rombach et al., 2022] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 10684–10695,

    2022

  13. [28]

    Countering misinformation via emotional response generation

    [Russo et al., 2023] Daniel Russo, Shane Kaszefski- Yaschuk, Jacopo Staiano, and Marco Guerini. Countering misinformation via emotional response generation. In Proceedings of the 2023 Conference on Empirical Meth- ods in Natural Language Processing, pages 11476–11492, December

    2023

  14. [29]

    Deep multimodal image- repurposing detection

    [Sabir et al., 2018] Ekraam Sabir, Wael AbdAlmageed, Yue Wu, and Prem Natarajan. Deep multimodal image- repurposing detection. In Proceedings of the 26th ACM in- ternational conference on Multimedia , pages 1337–1345,

    2018

  15. [32]

    Position: Will we run out of data? lim- its of LLM scaling based on human-generated data

    [Villalobos et al., 2024] Pablo Villalobos, Anson Ho, Jaime Sevilla, Tamay Besiroglu, Lennart Heim, and Marius Hobbhahn. Position: Will we run out of data? lim- its of LLM scaling based on human-generated data. In Forty-first International Conference on Machine Learning, ICML 2024,

    2024

  16. [33]

    Wong, and Lidia S

    [Wu et al., 2023] Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Derek F. Wong, and Lidia S. Chao. A sur- vey on llm-generated text detection: Necessity, methods, and future directions. CoRR, abs/2310.14724,

    2023 arXiv

  17. [34]

    Fake news in sheep’s clothing: Robust fake news detection against llm-empowered style attacks

    [Wu et al., 2024] Jiaying Wu, Jiafeng Guo, and Bryan Hooi. Fake news in sheep’s clothing: Robust fake news detection against llm-empowered style attacks. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3367–3378,

    2024

  18. [35]

    Wagner, Danqi Chen, and Prateek Mittal

    [Xiang et al., 2024] Chong Xiang, Tong Wu, Zexuan Zhong, David A. Wagner, Danqi Chen, and Prateek Mittal. Cer- tifiably robust RAG against retrieval corruption. CoRR, abs/2405.15556,

    2024

  19. [36]

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

    [Yao et al., 2023] 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 models. InProceedings of the 46th International ACM SIGIR Conference on Research and Developme...

    2023

  20. [37]

    Don’t take this out of con- text!: On the need for contextual models and evaluations for stylistic rewriting

    [Yerukola et al., 2023] Akhila Yerukola, Xuhui Zhou, Eliza- beth Clark, and Maarten Sap. Don’t take this out of con- text!: On the need for contextual models and evaluations for stylistic rewriting. In Proceedings of the 2023 Confer- ence on Empirical Methods in Natural Langua...

    2023

  21. [38]

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

    [Yuan et al., 2023] Xin Yuan, Jie Guo, Weidong Qiu, Zheng Huang, and Shujun Li. Support or refute: Analyzing the stance of evidence to detect out-of-context mis- and disin- formation. In Proceedings of the 2023 Conference on Em- pirical Methods in Natural Language Processing ,...

    2023

  22. [39]

    Instruction tuning for large language models: A survey

    [Zhang et al., 2023b] Shengyu Zhang, Linfeng Dong, Xi- aoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Fei Wu, and Guoyin Wang. Instruction tuning for large language models: A survey. CoRR, abs/2308.10792,

  23. [40]

    Detecting out-of-context multi- modal misinformation with interpretable neural-symbolic model

    [Zhang et al., 2023c] Yizhou Zhang, Loc Trinh, Defu Cao, Zijun Cui, and Yan Liu. Detecting out-of-context multi- modal misinformation with interpretable neural-symbolic model. CoRR, abs/2304.07633, 2023

    2023 arXiv

  24. [2015]

    Evidence-based factual error correction

    [Thorne and Vlachos, 2021] James Thorne and Andreas Vlachos. Evidence-based factual error correction. In Pro- ceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Vol- u...

    2021

  25. [2017]

    Survey of hallucina- tion in natural language generation

    [Ji et al., 2023] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucina- tion in natural language generation. ACM Comput. Surv., 55(12):248:1–248:38,

    2023

  26. [2018]

    Very deep convolutional networks for large-scale image recognition

    [Simonyan and Zisserman, 2015] Karen Simonyan and An- drew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9,...

    2015

  27. [2019]

    Hallucinated but factual! inspecting the factuality of hallu- cinations in abstractive summarization

    [Cao et al., 2022] Meng Cao, Yue Dong, and Jackie Cheung. Hallucinated but factual! inspecting the factuality of hallu- cinations in abstractive summarization. In Proceedings of the 60th Annual Meeting of the Association for Computa- tional Linguistics (Volume 1: Long Papers) ...

    2022

  28. [2020]

    Data scarcity, robustness and extreme multi- label classification

    [Babbar and Sch¨olkopf, 2019] Rohit Babbar and Bernhard Sch¨olkopf. Data scarcity, robustness and extreme multi- label classification. Machine Learning , 108(8):1329– 1351,

    2019

  29. [2021]

    Generating label cohesive and well-formed adversarial claims

    [Atanasova et al., 2020] Pepa Atanasova, Dustin Wright, and Isabelle Augenstein. Generating label cohesive and well-formed adversarial claims. In Proceedings of the 2020 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP), pages 3168–3177, November

    2020

  30. [2022]

    Catching out-of-context misinforma- tion with self-supervised learning.CoRR, abs/2101.06278,

    [Aneja et al., 2021] Shivangi Aneja, Christoph Bregler, and Matthias Nießner. Catching out-of-context misinforma- tion with self-supervised learning.CoRR, abs/2101.06278,

    2021 arXiv

  31. [2023]

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

    [Abdelnabi et al., 2022] 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

  32. [2024]

    [Dai et al., 2023] Wenliang Dai, Junnan Li, Dongxu Li, An- thony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. Instruct- BLIP: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Informa- ti...

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.