<p>Diffusion models have recently achieved impressive results in generative tasks, extending their capabilities from continuous data to discrete text. However, these models still face challenges when it comes to accurately capturing discrete entity information, especially in tasks like named entity recognition (NER). To address this, we propose MDNER, a novel NER model that integrates diffusion models with a Machine Reading Comprehension (MRC) framework. Unlike DIFFUSIONNER, which relies solely on diffusion processes without semantic priors, MDNER introduces MRC-guided denoising to explicitly align label semantics and contextual representations during boundary refinement. During training, MDNER concatenates the input text with predefined queries and processes them through BERT to obtain enhanced contextual representations. It then gradually adds noise to the golden entity boundaries and trains the model to recover these boundaries through a denoising process. In inference, the process starts by sampling noisy entity boundaries from a Gaussian distribution and inputs them with the text into MDNER. MDNER iteratively refines these noisy boundaries using the learned denoising process. By reformulating the NER task as a denoising diffusion process on entity boundaries, MDNER leverages the shared denoising objectives of diffusion models and pre-trained language model (PLM) to generate entity spans from noise. Experiments demonstrate that MDNER achieves state-of-the-art results across flat and nested NER datasets, with improvements of +0.18% and +0.78% in F1 scores on CoNLL2003 and OntoNotes5.0, respectively. For nested NER datasets such as ACE2005, it achieves gains of +1.66% in F1 scores. These results highlight MDNER’s superior performance in complex entity recognition scenarios.</p>