Text classification, where each input text is assigned to a single category, is a fundamental task in natural language processing. In this paper, we propose a novel framework that combines BERT-based contextual embeddings with a reinforcement learning technique known as group relative policy optimization (GRPO). To enrich input representations, we employ prompt-based text generation using GPT models. It helps create diverse paraphrased variants that are fused with the original embeddings. It is then used to train a classifier using the GRPO, which improves stability and reduces reliance on conventional supervised objectives. We conduct experiments on the baseline BERT model and perform an ablation study of our proposed pipeline using a financial fraud classification dataset. It shows that the integration of generative augmentation and policy-based learning leads to improved classification performance and training efficacy.