The field of natural language processing is witnessing a significant shift towards diffusion-based language models, which offer a promising alternative to traditional autoregressive models. Recent research has focused on addressing the limitations of diffusion models, such as information loss and premature commitment, and exploring new inference strategies to improve their efficiency and effectiveness. Notably, the development of novel decoding frameworks and cache eviction methods has led to substantial improvements in generation quality and speed. Furthermore, the connection between diffusion models and recurrent-depth models has been investigated, revealing new opportunities for parallelization and acceleration. Overall, the field is moving towards more efficient, flexible, and powerful language models. Noteworthy papers include: Latent Refinement Decoding, which introduces a two-stage framework for refining belief states and improving generation quality. On the Reasoning Abilities of Masked Diffusion Language Models, which characterizes the reasoning capabilities of masked diffusion models and demonstrates their equivalence to well-established reasoning frameworks.