The Paradigm Shift: A Comprehensive Survey on Large Vision Language Models for Multimodal Fake News Detection

TL;DR

The paper surveys the shift to large vision-language models for multimodal fake news detection, detailing a triadic reasoning framework that combines authenticity, cross-modal consistency, and manipulative intent. It introduces a three-branch taxonomy (Parameter-Frozen, Parameter-Tuning, Reasoning) and provides a structured framework for evaluating LVLM-based MFND methods against a broad set of benchmarks and datasets. Key contributions include a unified LVLM-centric perspective, a comprehensive dataset taxonomy, and a discussion of practical challenges such as hallucination, robustness, and real-time deployment, with concrete directions like causal reasoning and knowledge-enhanced adaptation. The work highlights the practical implications for deploying trustworthy, scalable, and interpretable MFND systems in real-world media environments.

Abstract

In recent years, the rapid evolution of large vision-language models (LVLMs) has driven a paradigm shift in multimodal fake news detection (MFND), transforming it from traditional feature-engineering approaches to unified, end-to-end multimodal reasoning frameworks. Early methods primarily relied on shallow fusion techniques to capture correlations between text and images, but they struggled with high-level semantic understanding and complex cross-modal interactions. The emergence of LVLMs has fundamentally changed this landscape by enabling joint modeling of vision and language with powerful representation learning, thereby enhancing the ability to detect misinformation that leverages both textual narratives and visual content. Despite these advances, the field lacks a systematic survey that traces this transition and consolidates recent developments. To address this gap, this paper provides a comprehensive review of MFND through the lens of LVLMs. We first present a historical perspective, mapping the evolution from conventional multimodal detection pipelines to foundation model-driven paradigms. Next, we establish a structured taxonomy covering model architectures, datasets, and performance benchmarks. Furthermore, we analyze the remaining technical challenges, including interpretability, temporal reasoning, and domain generalization. Finally, we outline future research directions to guide the next stage of this paradigm shift. To the best of our knowledge, this is the first comprehensive survey to systematically document and analyze the transformative role of LVLMs in combating multimodal fake news. The summary of existing methods mentioned is in our Github: \href{https://github.com/Tan-YiLong/Overview-of-Fake-News-Detection}{https://github.com/Tan-YiLong/Overview-of-Fake-News-Detection}.

Figures (6)

• Figure 1: A chronological overview of representative LVLMs is presented, highlighting the rapid growth of this field.
• Figure 2: Taxonomy of Multimodal Fake News Detection and Veracity Reasoning. We systematically categorize multimodal fake news detection methods according to their parameter adaptation paradigms, including full fine-tuning, parameter-efficient tuning, prompting-based inference, and agent-based reasoning. This taxonomy provides a structured overview of state-of-the-art approaches and clarifies how different parameter interaction strategies are leveraged for multimodal veracity assessment.
• Figure 3: Zero-shot learning framework for multimodal fake news detection. The architecture integrates direct prediction and reasoning mechanisms to process text and image inputs. It employs a multi-query generation module to formulate news-related queries based on the input's title and keywords. These queries are filtered for topical relevance and used to extract evidence from external sources. If external knowledge is deemed necessary, the system verifies the authenticity of both text and image inputs. The extracted evidence, along with the initial predictions, is refined through a reasoning process to produce the final output. This end-to-end approach enables the detection of fake news without prior training on labeled data, leveraging the complementary strengths of direct reasoning and external evidence integration xuan2024lemma.
• Figure 4: Parameter-efficient tuning architecture for multimodal fake news detection. The framework operates in two stages: global knowledge learning and complementary knowledge fusion. In the first stage, a student model with 7 billion parameters is trained alongside two adapters, processing visual and textual inputs to identify relevant entities and search options. The second stage integrates the student model with the adapters to enhance reasoning. Knowledge acquisition is achieved through two vision-language models, Qwen2-VL and InternVL, which provide predictions and rationales for the input image. Multi-teacher knowledge distillation follows, where the student model and adapters align their predictions and rationales, facilitated by LoRA and DPO techniques to refine the model's outputs zeng2024multimodal.
• Figure 5: Timeline of multimodal fake news detection datasets.