LLM-Enhanced Misinformation Detection and Explainable harms Tracking (Jing Ma et al.)

Misinformation on social media, such as rumors and fake news, poses significant challenges to information integrity and public trust. Recent advances have explored how temporal dynamics and structural diffusion patterns can reveal hidden signals of misinformation. Neural architectures such as recurrent models, recursive tree-based networks with attention, and Transformer-based methods provide powerful ways to capture both propagation sequences and complex user interactions.

At the same time, domain-specific misinformation, such as health-related rumors, requires methods that go beyond open-domain features. Contrastive learning, prompt-based zero-shot detection, and cross-lingual adaptation have emerged as promising approaches for handling low-resource and multilingual scenarios. Furthermore, explainability is becoming crucial: sentence-level evidence retrieval, entailment-based reasoning, and weakly supervised fine-grained detection enable not only accurate classification but also transparent justifications of model decisions.

Beyond textual misinformation, harmful memes have emerged as a new and particularly difficult challenge. Harmful memes often rely on implicit and metaphorical meanings conveyed through the interaction of images and text. LLM-based multimodal reasoning frameworks use abductive inference to uncover hidden intent and fine-tuned generative models for detection.

Looking forward, LLMs open new opportunities for misinformation research. By combining prompt learning, retrieval-augmented generation, and fine-grained supervision, LLMs can be leveraged to detect, explain, and adapt to misinformation in real time across domains, languages, and platforms.


References:


For further information on this research topic, please contact Prof. Jing Ma.