Large Language Models (LLMs) often underperform on know-ledge-intensive tasks due to insufficient domain-specific information and a tendency to hallucinate. While knowledge graphs (KGs) offer structured and reliable information, integrating them into LLMs poses challenges related to semantic misalignment and inconsistent usage during text generation. To address these issues, we propose EntiFA, a fine-grained adaptation framework that dynamically integrates entity embeddings from KGs with the contextual representations of LLMs. EntiFA features a dual-tower architecture to align semantic spaces and a dynamic embedding fusion mechanism that operates throughout the generation process. By employing a parameter-efficient adaptation layer, EntiFA enables consistent and effective knowledge integration without compromising the model’s original capabilities. Extensive experiments across multiple benchmarks show that EntiFA significantly outperforms existing methods in knowledge-intensive tasks, achieving superior accuracy and efficiency.