Detecting Out-of-Context Misinformation via Multi-Agent and Multi-Grained Retrieval

1. Vague Motivation: The motivation presented in this paper lacks clarity. The authors claim: "We developed the first multi-agent OOC detection system that integrates multi-granularity information, mirroring the real-world collaboration among human experts." However, as far as I know, human experts responsible for misinformation detection typically make independent judgments and reach a final consensus through discussion, rather than each expert being responsible for only a single part of the detection process. In contrast, MACAW assigns each agent responsibility for only one specific step of the detection process. This distinction seems misaligned with how human experts collaborate.
   https://www.politifact.com/article/2018/feb/12/principles-truth-o-meter-politifacts-methodology-i/

2. Lack of Literature Review: The paper overlooks some relevant works that focus on enhancing explainability, which should be acknowledged. This includes, but is not limited to:

   • Interpretable Multimodal Misinformation Detection with Logic Reasoning (Findings of ACL 2023)
   • Interpretable Detection of Out-of-Context Misinformation with Neural-Symbolic-Enhanced Large Multimodal Model (Findings of ACL 2023)
   • Interpretable Multimodal Out-of-context Detection with Soft Logic Regularization (ICASSP 2024)

3. Insufficient Details on Data Storage: The paper lacks sufficient detail on how the data storage is constructed. What are the data sources? Does it retrieve information from the internet? If so, have the authors accounted for potential time leakage during retrieval? For instance, some misinformation may have already been flagged as fake news by the time it's retrieved from the web. This could introduce unintended biases, particularly when handling older misinformation cases.

4. Incomplete Experimental Reports: The experimental results are incomplete. Table 1 only shows results on the Merged/Balance subset of the NewsCLIPpings dataset. What about other subsets, such as Text-Image, Text-Text, Person-Matching, and Scene-Matching? A broader range of experiments would give a clearer picture of the strengths and weaknesses of MACAW framework.

5. Unreasonable Explainability Analysis: Section 4.4.2, Explainability Analysis, is insufficient. The paper compares different backbone models within MACAW framework, but it should instead provide a comparison with other OOC detection approaches in terms of explainability. The current comparison only shows how GPT-4o performs better than other models, which is not enough to assess the explainability of the system. Additionally, incorporating more qualitative analysis in this section would further strengthen the evaluation.

W1) The authors utilize the NewsCLIPpings dataset, which includes ‘out-of-context’ samples that are ‘synthetic’ or algorithmically created by mismatching the original image-text pairs. However, strong performance on ‘synthetic’ data does not guarantee similar results on real-world data. The absence of evaluation against real-world out-of-context misinformation is a notable limitation.

W2) There is no comparison with the current SotA method on NewsCLIPpings such as the “Attentive Intermediate Transformer Representations” (AITR) model [1] which achieves 93.3% on NewsCLIPpings. Moreover, the authors of [1] express concerns regarding the use of the NewsCLIPpings dataset, noting how models can achieve high performance on it by relying on superficial patterns rather than factuality.

W3) While the ablation study demonstrates that each component of the proposed method is important, Table 3 reveals the significance of using the more powerful GPT-4o-latest (92.7%) over GPT-4o-mini, which achieves only 84.6%. This performance is surpassed by previous methods, such as CCN (84.7%) and SNIFFER (88.4%). Consequently, this somewhat diminishes the significance of the proposed method, suggesting that the observed performance improvement is primarily attributable to the more powerful LVLM rather than the method itself.

W4) Table 4 presents the ranking of each model’s explanations and logical consistency. However, this does not necessarily indicate the quality of the explanations, just that they are preferred over the explanations of other methods.

W5) The presentation of competing methods is somewhat superficial, lacking an in-depth discussion of how the proposed method distinguishes itself from other LVLM-based approaches, such as SNIFFER or [2, 3].

References [1] Papadopoulos, S. I., Koutlis, C., Papadopoulos, S., & Petrantonakis, P. C. (2024). Similarity over Factuality: Are we making progress on multimodal out-of-context misinformation detection?. arXiv preprint arXiv:2407.13488. [2] Tahmasebi, S., Müller-Budack, E., & Ewerth, R. (2024, October) Multimodal Misinformation Detection using Large Vision-Language Models. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management (pp. 2189-2199). [3] Geng, J., Kementchedjhieva, Y., Nakov, P., & Gurevych, I. (2024). Multimodal Large Language Models to Support Real-World Fact-Checking. arXiv preprint arXiv:2403.03627.

1. The core of this work should focus on the accuracy of image-text matching in the evidence stored within the module and the quantity of evidence it contains. Moreover, to ensure high accuracy in OOC detection, the evidence storage module may need constant updates; otherwise, maintaining a high detection accuracy for OOC over time may be challenging. Consequently, this might limit the work’s impact on future OOC research.
2. In the example of image-text detection shown in Figure 1, the image depicts Cameron wearing fall-winter attire, with the caption stating that Cameron left the High Court on June 14. This discrepancy is quite evident, appearing solvable without extensive background data. Figure 2 revisits this example but it still relies on empirical information for news detection.