Dynamic Analysis and Adaptive Discriminator for Fake News Detection

TL;DR

DAAD introduces a dynamic, two-stage framework for multimodal fake news detection that couples Monte Carlo Tree Search–driven prompt optimization of LLM comments with an Adaptive Discriminator ensemble. The system uses a MemoryBank, Batchprompt, and Resampling to stabilize prompt search, while four specialized discriminators (ReLU, Frequency Domain, Logical, Semantic) operate under a Soft Router to adaptively fuse cues. Across Weibo, Weibo-21, and GossipCop, DAAD delivers state-of-the-art results, highlighting the value of domain-specific prompts and flexible discriminator routing for robust MFND. The approach advances practical fake news detection by combining flexible language-model guidance with a modular, data-driven discrimination strategy, and it includes a public code release to facilitate adoption and replication.

Abstract

In current web environment, fake news spreads rapidly across online social networks, posing serious threats to society. Existing multimodal fake news detection methods can generally be classified into knowledge-based and semantic-based approaches. However, these methods are heavily rely on human expertise and feedback, lacking flexibility. To address this challenge, we propose a Dynamic Analysis and Adaptive Discriminator (DAAD) approach for fake news detection. For knowledge-based methods, we introduce the Monte Carlo Tree Search algorithm to leverage the self-reflective capabilities of large language models (LLMs) for prompt optimization, providing richer, domain-specific details and guidance to the LLMs, while enabling more flexible integration of LLM comment on news content. For semantic-based methods, we define four typical deceit patterns: emotional exaggeration, logical inconsistency, image manipulation, and semantic inconsistency, to reveal the mechanisms behind fake news creation. To detect these patterns, we carefully design four discriminators and expand them in depth and breadth, using the soft-routing mechanism to explore optimal detection models. Experimental results on three real-world datasets demonstrate the superiority of our approach. The code will be available at: https://github.com/SuXinqi/DAAD.

Figures (5)

• Figure 1: A comparison is made between our optimized prompt vs original prompt using LLMs, such as Chinese LLaMA, for news authenticity classification. The optimized prompt offers richer and domain-specific guidance (highlighted in red), resulting in more accurate comments and predictions (highlighted in gray), such as exaggerations (highlighted in yellow), details (highlighted in blue).
• Figure 2: Some examples illustrate the reasons for fake news: (a) the text exhibits clear emotional bias; (b) the text contains exaggerated emotional bias and defies common sense; (c) the image has been manipulated; and (d) there is a semantic inconsistency between the image and the text.
• Figure 3: The architecture of our proposed DAAD for fake news detection consists of two main steps: comment on the news by LLMs and adaptive discriminator modeling. First, the news is processed by a prompt-optimized LLM to generate comments, which are then aggregated by the Adaptive Comment Aggregation module. Next, the Adaptive Discriminator module analyzes the cues with various discriminators. Finally, the model outputs the probability that the news is fake.
• Figure 4: (a) An example of a state transition. Given a prompt ($s_t$), Step 1 is responsible for collecting errors, Step 2 provides modification suggestions ($a_t$) through a LLM, and Step 3 generates the new prompt ($s_{t+1}$) through other LLM. (b) An example of Batch Prompt. (3) The MemoryBank collects historical errors and summarizes them into global guidance, which is fed back into the state transition process.
• Figure 5: Visualization of routing paths in some fake news examples. The texts are translated from Chinese to English.