Optimization Techniques in Large Language Models for News Report Generation

This paper aims to explore optimization techniques in Large Language Models (LLMs) for automated news report generation, utilizing methods such as Retrieval-Augmented Generation (RAG), Zero-Shot Learning, ICL (In Context Learning), and reAct agents. Using these techniques, the system collects real-time information based on the user context and the images from the internet, to improve factual accuracy and enhance reliability. Various LLMs such as Llama 3.1, Llama 3.2, Qwen 2.5, Gemma, LLaVa, Aya were utilized in the system to generate the News Reports. Among these, Llama3.1 demonstrated the best performance with Context Precision as 92%, Answer Relevancy as 90.84 % and Faithfulness as 12.5%, closely followed by Qwen 2.5, which showed a notable improvement in Faithfulness as 43%. Fine-tuning the models can improve the various metrics such as Context Precision, Recall, Answer Relevancy, BLEU and ROUGE Scores as well as improve the adoption to various writing styles as seen in News Reports.