Sequential recommendation (SR) aims to model users’ dynamic interests over time and predict their next actions. However, existing sequential recommendation models based on conditionally guided diffusion still face two key challenges in long-term sequence modeling. First, excessive noise injection may disrupt the structure of target embeddings, thereby affecting recommendation accuracy and stability. Second, the high computational complexity of processing long sequences limits model efficiency. To tackle these challenges, we propose the Efficient Extraction of Long-Term Sequential Signals for Guiding Diffusion-Based Recommendation (ELS-GDR). We design ELS-GDR to effectively extract long-term sequential signals and enhance diffusion-based recommendation. Specifically, we incorporate an L2 normalized linear attention mechanism into the long-term sequential recommendation framework, significantly reducing computational complexity while capturing users’ long-term behavioral patterns more effectively. We further introduce a bias noise strategy to improve the model’s robustness against noisy interactions. Finally, we integrate a classifier-free guided diffusion mechanism to optimize the reverse denoising process, enhancing recommendation accuracy and personalization. Extensive experiments on three public datasets demonstrate that our method consistently outperforms existing approaches in both recommendation performance and computational efficiency.