Hybrid Transformer–LLM Framework with Factual Reliability Features for Fake News Detection

The proliferation of fake news undermines public trust, destabilizes societies, and erodes democratic processes. In this work, we propose a hybrid transformer-LLM framework that integrates transformer-based architectures with credibility-aware features generated by a state-of-the-art Large Language Model (LLM). Our approach employs two complementary processing streams: (1) a fine-tuned BERT model trained on the original dataset, and (2) a second BERT model applied to features derived from LLM via a carefully engineered zero-shot prompting template. This template is designed to elicit semantic and authenticity-related descriptors that capture nuanced cues of misinformation. The outputs from both BERT models are fused to produce the final prediction, enabling the system to leverage both raw linguistic patterns and enriched LLM-generated context. Experimental results on real-world datasets demonstrate that our framework achieves a 2% improvement in accuracy and a 4% improvement in precision over state-of-the-art baselines, validating the effectiveness of integrating LLM prompting with transformer-based detection pipelines.