Group Related Policy Optimization-Enhanced Framework for Text-Based Fake News Detection

The swift dissemination of disinformation is a quickly emerging obstacle in today’s digital era, and social media websites are accelerating the propagation of false content, which is frequently backed up by AI-created content that is hard to distinguish from actual information. The aim of this research is to improve the effectiveness and precision of a fake news detection system by assessing the effect of Group Relative Policy Optimization (GRPO) on the performance metrics of two prominent large language models, specifically Qwen 2.5-3B-Instruct and LLaMA 3.2. The two models are assessed independently with an equalized dataset containing original and generated news to allow easy comparison of their performance outcomes with and without the application of GRPO. The method employs a twin reward function design that prioritizes response succinctness as well as classification accuracy. Optimization techniques such as quantization and LoRA are employed to counteract the computational burden. The evaluation results indicate that GRPO achieves: impressive accuracy gains, from 73% to 87% for Qwen, and from 59% to 93% for LLaMA, improved F1-scores and balanced classification for both classes (Fake/Real), a sharp reduction in misclassifications and false positives, and considerable improvement in ROC and Precision-Recall curves, in terms of increased AUC values. This study represents the first recorded implementation of GRPO for fake news detection and provides a detailed examination of its implications on various language models. The results highlight the power of GRPO as an improved training methodology that enhances model credibility and efficiency and paves the way for novel mechanisms to tackle misinformation in dynamic media landscapes.