CURE: Context-aware Uncertainty-adaptive Rebalancing Experts for Multimodal Fake News Detection

ACL ARR 2026 January Submission5470 Authors

Keywords: Multimodal Fake News Detection, Modality Imbalance, Retrieval Augmentation, Mixture of Experts

Abstract: The proliferation of fake news on social media motivates the development of automatic Multimodal Fake News Detection. While most existing methods focus on various fusion strategies, they largely overlook the heterogeneous modality imbalance issue (inter-modal information disparities and noise interference) in real-world scenarios, which hinders effective fusion through biased feature representations. To address this, we propose a novel Context-aware Uncertainty-adaptive Rebalancing Experts (CURE) framework for multimodal fake news detection. First, to bridge inter-modal information disparities, a Retrieval-Augmented Context Prompter (RACP) module retrieves similar instances and distills them into dynamic prompts, enhancing the features of each modality by providing supplementary context. Second, to mitigate noise interference, an Uncertainty-Adaptive Rebalancing Experts (UARE) module first quantifies feature-level noise via uncertainty modeling. An uncertainty-adaptive routing mechanism then achieves robust modality rebalancing by adaptively down-weighting features with high uncertainty. Extensive experiments are conducted on three real-world datasets spanning two languages, demonstrating significant performance improvements of our method. The code is available at https://anonymous.4open.science/r/CURE1.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: rumor/misinformation detection, multimodal applications, multimodality

Contribution Types: NLP engineering experiment

Languages Studied: Chinese, English

Submission Number: 5470