Automated text summarization is increasingly crucial due to the vast amounts of information generated daily across various domains, necessitating the development of models that can produce concise and coherent summaries. The integration of reinforcement learning with an enhanced Mistral LLM introduces a novel approach that dynamically refines summarization tasks through iterative feedback, resulting in improved quality and relevance. The proposed model utilizes architectural modifications, including multi-head attention and gated mechanisms, alongside a propagation feedback mechanism within the reinforcement learning framework, to optimize the generation of summaries. Experimental results reveal that the enhanced model surpasses traditional and contemporary baselines in key evaluation metrics such as ROUGE, BLEU, and METEOR scores, indicating a higher degree of fluency, coherence, and informativeness. The qualitative assessment further corroborates the model's ability to generate human-like summaries, making it a valuable tool for applications requiring the distillation of large volumes of text into essential information. Additionally, the research provides insights into the potential use of reinforcement learning for enhancing LLMs in various text generation tasks, demonstrating significant implications for future advancements in text summarization and related areas.