Hierarchical Progressive Alignment: A Cognitive-inspired Framework for Short-Video Fake News Detection

ACL ARR 2026 January Submission1367 Authors

Keywords: Short-video fake news detection, Multimodal misinformation, Social media content verification

Abstract: Fake News Detection is a very important but challenging task that has attracted the attention of both the field of natural language processing and multimedia computing. Most existing approaches to fake news detection on short-video platforms adopt static cross-modal fusion, directly combining visual content with auxiliary modalities such as text and audio for classification. While effective in some cases, this paradigm is sensitive to modality-specific noise and tends to overfit superficial cross-modal correlations, particularly on easy samples, which can undermine robustness. To address these issues, we introduce HPA, a diffusion-driven Hierarchical Progressive Alignment framework that performs adaptive computation. HPA begins with a lightweight authenticity predictor that produces an initial decision along with a confidence estimate, and selectively routes uncertain samples to subsequent stages. For these samples, a diffusion-based opinion evolution module iteratively denoises and reconstructs modality-specific semantic opinions, encouraging alignment in a shared latent space through a reconstruction objective. A fine-grained attribution module then refines the final prediction and provides cues associated with potential manipulation. Experimental results on FakeSV and FakeTT show that HPA achieves consistent improvements over strong baselines and generalizes well across datasets.

Paper Type: Long

Research Area: Multimodality and Language Grounding to Vision, Robotics and Beyond

Research Area Keywords: Multimodaility and Language Grounding to Vision,Robotics and Beyond

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 1367