This research work investigates the limitations of current Large Language Models (LLMs) in maintaining contextual information over extended text inputs. While they have been efficacious with various natural language processing (NLP) tasks, standard attention mechanisms are weak with long-range dependencies, resulting in decreased coherence and relevance in output generations. To meet this challenge, an adaptive attention mechanism (AAM) is proposed, which dynamically adapts the attention weights by contextual relevance so that the model can concentrate on the most appropriate tokens when processing longer text. This experimental configuration compares the performance of the projected work against a set of existing literature. This paper suggests a new long-text narrative dataset to evaluate the model's ability to deal with complex contexts. Findings indicate that the proposed adaptive attention mechanism significantly improves performance metrics with a 4% improvement in F1 score on the SQuAD dataset and 3.6% in ROUGE scores for summarization tasks. The findings suggest that adaptive attention mechanisms can benefit LLMs’ capability by an impressive margin, opening the way for more contextual and coherent text generation in actual situations. Ultimately, this work paves the way to pushing the state-of-the-art for NLP for-ward by fulfilling the need for improved contextual understanding in LLMs.