acceptodds
Under review as a conference paper at ICLR 2027

Evidential Multi-view Uncertainty Learning for Multimodal Fake News Detection

Abstract

Multimodal fake news detection benefits from complementary textual and visual cues; however, existing methods often treat local text tokens and image patches as equally reliable. Consequently, intra-modal noise may propagate through cross-modal interaction and impair final decisions. To overcome this challenge, we present Evidential Multi-view Uncertainty Learning (EMUL), a unified framework that seamlessly integrates uncertainty as a continuous regulatory signal across representation learning, semantic alignment, and decision fusion. Rather than processing components in isolation, EMUL first estimates token- and patch-level intra-modal uncertainty to perform confidence-weighted aggregation for fine-grained noise suppression. These estimated confidences directly condition cross-modal semantic alignment, reducing the distribution gap between reliable local modality representations and global image-text semantics while down-weighting uncertain samples. Ultimately, at the decision stage, this uncertainty-awareness enables EMUL to jointly calibrate the text, image, and fused views using intra-modal and evidential uncertainty prior to multi-view evidence fusion. Extensive experiments on three public benchmarks demonstrate that EMUL achieves state-of-the-art performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Uncertainty-Guided, Multi-view Evidential, Multimodal Fake News Detection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HM4fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors observe that, in multimodal fake news detection, directly performing cross-modal interaction and fusion may propagate noise from individual modalities into the fused representation, thereby impairing the final prediction. To address this issue, they propose an Evidential Multi-view Uncertainty Learning (EMUL) framework that uses uncertainty as a regulatory signal throughout the detection pipeline. Specifically, EMUL consists of three stages: uncertainty-aware multi-view representation learning, which estimates token- and patch-level uncertainty and performs confidence-weighted aggregation to obtain more reliable modality representations; uncertainty-guided semantic alignment, which uses the estimated uncertainty to modulate the alignment between local modality representations and global image-text semantics; and reliability-aware evidential fusion, which jointly considers intra-modal and evidential uncertainty to calibrate and fuse the text, image, and fused views. Compared with a range of representative baselines, EMUL achieves the best performance on Weibo and Weibo-21, while remaining competitive on GossipCop.

Critical strengths
  1. EMUL provides a coherent three-stage framework for suppressing unreliable intra-modal information before and during multimodal interaction and fusion. Specifically, the framework first performs uncertainty-aware multi-view representation learning to estimate token- and patch-level intra-modal uncertainty and suppress unreliable local units through confidence-weighted aggregation, then applies uncertainty-guided semantic alignment, and finally conducts reliability-aware evidential fusion to produce the final prediction. This design directly targets a limitation identified in prior multimodal fake news detection methods, where encoded local units are often treated as reliable inputs and token-/patch-level noise is insufficiently controlled before cross-modal interaction and decision fusion.
  2. EMUL demonstrates strong empirical performance across multiple benchmarks, and the ablation studies provide evidence for the usefulness of uncertainty-aware representation learning. The proposed method achieves the best performance on Weibo and Weibo-21 and remains competitive on GossipCop. In addition, removing TCUE consistently degrades performance, while the local-corruption experiments further show that uncertainty-aware representation learning improves robustness under synthetic textual and visual perturbations. These results are consistent with the intended effect of suppressing unreliable local information and suggest that explicitly modeling local reliability before multimodal interaction is a useful design choice that may provide insights for future multimodal learning methods.
Critical weaknesses

1.The motivation regarding the propagation of intra-modal noise could be better substantiated. The paper argues that unreliable tokens or patches within each modality may propagate through cross-modal interaction and fusion and eventually impair the final prediction. This is a reasonable motivation, and the local-corruption experiments provide evidence that explicitly suppressing unreliable local information can improve performance. However, the current analysis does not systematically demonstrate how intra-modal noise is propagated or amplified by existing multimodal fusion methods. In particular, it remains unclear to what extent prior methods are unable to implicitly suppress such noise or uncertainty during end-to-end training. A more systematic analysis of how noisy local information affects representations before and after multimodal interaction, or how representative baselines respond to controlled intra-modal corruption, would better support this central motivation. Otherwise, the corresponding claims about noise propagation or amplification may need to be stated more conservatively.

2.The novelty of the individual components appears relatively modest, with the main contribution lying in their integration into a unified uncertainty-aware framework. TCUE, UGSA, and RAEF are built on established ideas such as variational uncertainty estimation, adversarial semantic alignment, and evidential/Dempster-Shafer fusion. The paper’s main novelty therefore seems to come from consistently propagating intra-modal uncertainty across representation learning, semantic alignment, and decision fusion, rather than from a fundamentally new mechanism in each component.

Questions for authors

I encourage the authors to address some of the following issues in the rebuttal and I would adjust the rating based on the results.

1.The paper motivates EMUL by arguing that intra-modal noise may propagate through cross-modal interaction and fusion and eventually impair the final prediction. Could the authors provide further evidence or analysis showing how such noise affects existing multimodal fusion methods? In particular, to what extent can conventional end-to-end training implicitly suppress unreliable information from individual modalities?

2.The paper does not specify the coefficient values used for the loss terms in the main experiments, such as λaux\lambda_{\mathrm{aux}}, λedl\lambda_{\mathrm{edl}}, λvar\lambda_{\mathrm{var}}, and λugsa\lambda_{\mathrm{ugsa}}. Could the authors clarify the exact values used? Please also clarify whether these coefficients are kept consistent across different datasets or selected separately for each dataset.

3.Could the authors provide experimental results on at least one additional multimodal fake news detection dataset beyond Weibo, Weibo-21, and GossipCop? This would help further assess the generalization of EMUL across different datasets.

Reviewer_Qj6jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes EMUL for multimodal fake news detection. The framework estimates token level and patch level uncertainties using latent Gaussian variance and contextual residuals. Then it uses the resulting confidence scores for representation aggregation and adversarial alignment with CLIP features. And then it discounts text, image, and fused view evidential opinions before Dempster Shafer fusion.

The experiments cover three datasets, with component ablations, local corruption tests, selective prediction, and additional calibration and error detection analyses. The application is relevant, but I have concerns about the formulation of the alignment module, the interpretation of local uncertainty, and the strength and reproducibility of the empirical evidence.

Critical strengths
  1. The paper addresses a meaningful problem, i.e., unreliable local content can affect multimodal representations before decision level fusion. Using local reliability throughout the pipeline is a reasonable research direction.

  2. The evaluation goes beyond classification accuracy. I appreciate the corruption experiments, the uniform versus uncertainty weighted alignment comparison, and the calibration and error awareness metrics in Appendix C.

  3. The reported classification results are promising on the two Chinese datasets. The appendix also appropriately acknowledges that improved error ranking does not imply uniformly improved probability calibration.

Critical weaknesses
  1. The same modal and cross modal discriminator objectives appear equivalent as shown in Appendix A1. Eqs (17)-(19) sum over both modalities. For the same modal discriminator, the positive examples are the global text and image representations, and the negative examples are the local text and image representations. For the cross modal discriminator, exchanging modality mm with its complementary modality merely permutes the same positive examples; the negative examples remain unchanged. Hence, both discriminators solve the same global versus local distribution discrimination problem, albeit with separate parameters. The equations do not establish distinct same modal and cross modal alignment tasks. Moreover, each discriminator receives an individual representation rather than a paired or modality conditioned input, so the objective does not directly constrain instance level semantic correspondence. If the implementation uses modality specific discriminators, conditional inputs, or additional pairing information, these details must be specified and the equations corrected. Otherwise, the claimed two level alignment mechanism is not supported by the formulation.

  2. The central local uncertainty estimates are insufficiently validated and may be confounded with learned gating. The variance and contextual residual in Eqs (5)-(7) are plausible heuristics, but neither is demonstrated to measure local corruption or task relevant unreliability. A contextual outlier can be an important entity, negation, or discriminative visual detail rather than noise. Likewise, variance learned through classification and KL regularization does not automatically represent uncertainty about the correctness or usefulness of a token. The Gaussian prior also affects the interpretation: at unit variance, Eq (5) yields approximately 0.632, independently of whether a local unit is corrupted. This is not inherently invalid, but it shows that the score’s absolute meaning depends on the representation scale and regularization.

  3. The evaluation does not sufficiently establish the claimed advantage over the closest alternatives. The paper discusses uncertainty aware fake news detection and trust discounted evidential learning, but does not experimentally compare against several of the closest approaches, including MM-ULN. Comparisons with compatible methods under matched encoders and supervision would better establish the contribution beyond existing uncertainty weighting and evidential fusion. The state of the art claim also needs qualification. On GossipCop, EMUL matches the best reported accuracy, but its fake-news F1 is 0.680 versus 0.698 for MIMoE-FND, and its fake-news recall is 0.578 versus 0.644. These are important differences on an imbalanced dataset. Finally, higher corrupted-input accuracy does not by itself establish smaller degradation. On GossipCop, the full model drops from 0.895 to 0.852 under 30% text corruption, whereas w/o TCUE drops from 0.883 to 0.850. Please report both absolute corrupted performance and degradation from clean performance, ideally across multiple corruption draws.

  4. The core TCUE operators T_a and Agg are not concretely defined. The paper also lacks sufficient detail on image-patch ordering, the fusion encoder, discriminator architectures, encoder freezing, and whether CLIP and its adapters are fixed alignment targets. These details are important for assessing both the method and its computational cost. Appendix B gives general dataset sizes rather than the exact retained train/validation/test counts and split identifiers. The learning rate, early stopping protocol, alignment warm up duration, final hyperparameter settings, and precise EDL loss and annealing schedule should also be provided. It is unclear which baseline results were reproduced under identical splits.

Questions for authors
  1. What distinguishes the training data or conditioning of the same modal and cross modal discriminators after summing over modalities?

  2. Do the local uncertainty scores identify the known corrupted tokens and patches, and how do they compare with ordinary learned attention?

  3. Are gradients propagated through the uncertainty-dependent alignment weights? How does stopping those gradients affect performance?

  4. What is the effect of including the fused-view opinion in addition to its two constituent modalities?

Reviewer_TjB4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EMUL, an uncertainty-aware framework for multimodal fake news detection. It integrates uncertainty modeling across three stages: token- and patch-level representation learning (TCUE), uncertainty-guided semantic alignment (UGSA), and reliability-aware evidential fusion (RAEF). Experiments on Weibo, Weibo-21, and GossipCop demonstrate competitive performance, with additional ablations, corruption robustness tests, and uncertainty calibration analyses.

Critical strengths
  1. The paper incorporates uncertainty into representation learning, semantic alignment, and decision fusion, rather than restricting uncertainty modeling to final predictions.

  2. Strong empirical performance. EMUL achieves the highest reported accuracy on Weibo and Weibo-21, with competitive results on GossipCop.

Critical weaknesses
  1. The paper does not clearly establish how its fine-grained uncertainty estimation, uncertainty-guided alignment, and evidential fusion advance existing uncertainty-aware learning paradigms. A more explicit discussion of the technical distinctions and unique benefits of the proposed mechanisms would strengthen the novelty claim.

  2. TCUE combines variational variance and contextual residuals to estimate token- and patch-level uncertainty. However, the assumption that contextual inconsistency indicates unreliability is not necessarily valid, e.g., semantically distinctive or important tokens and image patches may also deviate from their surrounding context. The paper lacks direct validation showing that its uncertainty scores correlate with actual local corruption, semantic irrelevance, or feature reliability. Consequently, the central claim that TCUE suppresses unreliable information remains insufficiently substantiated.

  3. RAEF discounts evidential opinions using uncertainty estimates from TCUE, but the reliability of this calibration depends on the validity of those estimates. Moreover, the three fused views are statistically dependent, since the fused view is derived from the same text and image representations. Applying Dempster–Shafer fusion without explicitly addressing such dependence raises concerns about evidence double-counting. The calibration results are also mixed. On GossipCop, Full RAEF has worse ECE (0.1239 vs. 0.0903) and Brier score (0.1306 vs. 0.1036) than ordinary fusion, challenging the claim of consistently improved calibration.

Questions for authors

The paper presents a well-organized framework and promising classification results. While the framework is well motivated, the paper does not sufficiently establish the distinctive methodological contributions of its uncertainty modeling and propagation mechanisms beyond closely related approaches. More importantly, the reliability of the proposed uncertainty estimates is not directly established, despite being central to all three stages of the framework. The dependence among evidential views and inconsistent calibration improvements further weaken the methodological justification. While the experimental results are encouraging, stronger mechanism-level validation and comparisons with simpler uncertainty-aware alternatives are needed to substantiate the paper's main contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.