acceptodds
Under review as a conference paper at ICLR 2027

GLED-Net: A Global-Local Evidence Decomposition Network for Multimodal Fake News Detection

Abstract

Multimodal fake news detection requires reasoning about both the overall consistency of an image–text pair and a compact set of textual cues that may carry discriminative information. We propose GLED-Net, a CLIP-based global–local evidence decomposition framework that organizes these views into a global narrative stream and a local cue stream. The global stream models the correspondence between the complete post and its image, whereas the local stream uses automatically extracted keywords to construct a complementary representation. Branch-level supervision encourages both streams to remain predictive, and a learned gate combines their outputs. Experiments on Weibo-16 and FNDC show strong performance relative to the evaluated supervised and zero-shot baselines. Same-backbone controls, component ablations, sensitivity analysis, and repeated runs distinguish the effect of the pretrained encoder from that of the proposed evidence organization. The results also show that the benefit of heuristic keyword cues and adaptive fusion is dataset-dependent, which motivates more adaptive and verifiable cue selection.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multimodal fake news detection, Evidence decomposition network, Dual-branch design

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_86T9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GLED-Net, a CLIP-based Global–Local Evidence Decomposition Network for multimodal fake news detection. The method decomposes multimodal information into two complementary branches: a global narrative branch, which models the correspondence between the full post text and the associated image, and a local cue branch, which uses keywords extracted from the text as a compact complementary view. Both branches are individually supervised, and their representations are combined through an adaptive gated fusion module. The model is primarily instantiated with Chinese-CLIP and evaluated on the Weibo-16 and FNDC datasets. The reported results show strong performance compared with the evaluated supervised and zero-shot baselines. The paper also provides component ablations, a frozen-CLIP comparison, hyperparameter sensitivity analysis, random-seed analysis on Weibo-16, a same-backbone control, and experiments with several alternative vision-language backbones.

Critical strengths
  1. The paper is clearly motivated and provides a reasonably careful empirical analysis of the proposed architecture. The global/local decomposition is intuitive for multimodal fake news detection, and the manuscript clearly explains the roles of the narrative and cue branches. I also appreciate that the authors provide several useful controlled experiments rather than relying solely on the final performance numbers. In particular, the component ablations examine the effects of CLIP fine-tuning, auxiliary supervision, the two branches, and adaptive fusion. The paper is also appropriately cautious in interpreting its results; for example, it explicitly notes that the local branch and adaptive fusion do not provide uniform gains on FNDC.
Critical weaknesses
  1. The methodological novelty appears relatively limited. The proposed global-local decomposition is intuitive and reasonable for multimodal fake news detection. However, several key components, including feature gating, discrepancy features, auxiliary branch supervision, and gated fusion, are established techniques. As acknowledged by the authors, the main contribution primarily lies in combining these components into a unified framework. While this task-oriented design has practical value, the overall architecture is relatively simple, and the methodological novelty appears somewhat limited. A clearer discussion of the distinctive advantages of the proposed design would help better highlight its technical contribution.
  2. The empirical evaluation is not yet sufficiently comprehensive to fully support the generality of the proposed framework. The paper provides experimental results on Weibo-16 and FNDC, along with several useful ablation studies. Nevertheless, the evaluation is limited to two relatively old Chinese datasets. Including an additional and more recent benchmark, such as Weibo-21, would help further assess the generalizability of the proposed framework. In addition, although five baselines are included in the main comparison, two are zero-shot MLLMs, leaving only three task-specific supervised baselines. Comparisons with more recent multimodal fake-news detection methods would provide a more comprehensive evaluation. Finally, the local branch relies on Jieba-based TF–IDF keyword extraction, while the current sensitivity analysis only considers the keyword budget KK. Investigating alternative keyword extraction and segmentation strategies would help better understand the robustness of the proposed local cue branch.
Questions for authors

1.The current experiments are restricted to Weibo-16 and FNDC. An additional dataset, such as Weibo-21 if compatible with the task formulation, would help demonstrate whether the proposed global-local decomposition generalizes beyond the two current benchmarks.

2.Two of the five methods in the main comparison are zero-shot MLLMs, leaving only three directly comparable supervised fake-news detectors. It would be helpful to include additional recent multimodal fake-news detection methods under comparable training settings.

3.The current implementation uses TF–IDF keyword extraction from Jieba, while the sensitivity study mainly varies the number of selected keywords KK. Have the authors considered alternative cue extraction approaches, entity-based extraction, alternative Chinese segmentation pipelines, or LLM-based keyword extraction? Such experiments would help clarify whether the observed benefit comes from meaningful cue selection or simply from providing an additional shortened textual representation.

4.The following minor issues do not affect my overall assessment or rating of the paper. In Figure 1, the purple cue feature in the cue branch does not appear to be clearly annotated with kk, and adding this notation may make the correspondence with the formulation easier to follow. It would also be helpful if Figure 1 explicitly indicated which components are fine-tuned and which are frozen, especially given that the paper additionally studies a frozen-CLIP setting. Finally, the font size and overall scale of Table 6 appear noticeably larger than those of the other tables; using a more consistent table style would improve the visual presentation of the manuscript.

Reviewer_y73sOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes GLED-Net, a CLIP-based framework for multimodal fake news detection, which models global image–text consistency and local keyword cues.

Critical strengths

The method is simple and easy to implement.

Critical weaknesses

This method offers limited novelty, its effectiveness varies across different datasets, and there is no in-depth analysis of the underlying reasons for this variation."Feature gating... are established neural components; our contribution lies in their task-specific organization into separately supervised global and local views for fake news detection."

The frequent use of negative constructions obscures the main argument."Figure 2: Illustrative FNDC cases. These examples provide a compact comparison between GLEDNet and FND-Bootstrap, but should not be interpreted as causal evidence or faithful explanations."

Some results reported in Table 2 appear unusual and require further clarification.

Questions for authors

No

Reviewer_DwLvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GLED-Net, a CLIP-based framework for multimodal fake news detection that decomposes image-text information into two complementary streams: a global narrative stream capturing overall image-text consistency and a local cue stream constructed from TF-IDF keywords. The framework employs branch-level supervision and adaptive gated fusion to combine the two representations. Experiments on Weibo-16 and FNDC demonstrate strong classification performance, accompanied by component ablations, sensitivity analyses, and repeated-run evaluations.

Critical strengths
  1. GLED-Net adopts a straightforward dual-branch architecture that explicitly separates global image-text consistency from local keyword-conditioned representations. The architecture is easy to understand and implement, with clearly defined responsibilities for each component.

  2. The paper explicitly examines the influence of CLIP fine-tuning through frozen-backbone and alternative-encoder comparisons. This is valuable because it helps distinguish improvements attributable to pretrained representations from those introduced by the proposed architecture, avoiding overstatement of the method's contribution.

  3. The authors acknowledge that the local cue stream and adaptive fusion do not consistently improve performance across datasets. This transparent interpretation of the experimental findings strengthens the credibility of the analysis and provides useful directions for future research.

Critical weaknesses
  1. While the global–local decomposition is intuitively motivated, the paper does not sufficiently distinguish its methodological contribution from existing multi-branch multimodal learning approaches. Specifically, the proposed branches primarily differ in whether the text input is the complete post or a TF-IDF keyword sequence, while employing similar feature interaction and classification mechanisms. It is suggested to clarify what distinctive modeling capability this decomposition introduces beyond conventional multi-view representation learning.

  2. Limited reliability of local evidence extraction. The local branch relies on TF-IDF to extract discriminative keywords, but lexical salience does not necessarily correspond to factual relevance or misinformation evidence. Important information may depend on entities, relations, or contextual expressions that cannot be preserved through isolated keyword extraction. Although the authors appropriately acknowledge that these keywords are not verified evidence, the paper lacks systematic validation of whether the selected cues capture meaningful claim-related information rather than superficial statistical correlations.

  3. The effectiveness of core components is inconsistent. The ablation results reveal that the local cue stream provides substantial improvements on Weibo-16 but no measurable benefit on FNDC, where removing it slightly increases F1 from 98.29% to 98.30%. Similarly, fixed mean fusion outperforms adaptive gated fusion on FNDC (98.49% versus 98.29%). These findings raise questions about the general applicability of the proposed global–local decomposition and adaptive fusion mechanisms.

  4. The FNDC preprocessing generates multiple image-text pairs from posts containing multiple images, followed by a random train/test split. Without explicitly ensuring post-level separation, samples sharing identical text and labels may appear across training and test sets, potentially inflating performance. Furthermore, the reproduced FND-Bootstrap baseline uses a randomly initialized image encoder because pretrained weights were unavailable, limiting the fairness of this comparison. Evaluation on only two Chinese datasets also leaves cross-domain and multilingual generalization insufficiently explored.

Questions for authors

The paper presents a clearly motivated framework and encouraging classification results. However, the distinctive methodological contribution of the global–local decomposition remains insufficiently established, and the reliance on heuristic keyword extraction raises concerns about the reliability of the proposed local evidence representation. More importantly, the inconsistent benefits of core components and potential data overlap in FNDC limit the strength of the empirical conclusions.The paper would benefit from stronger methodological comparisons, more principled evaluation of local evidence quality, and verification of dataset-level independence.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.