The field of vision-language models is currently moving towards improving the alignment between visual and language modalities, with a focus on enhancing the robustness and fine-grained understanding of these models. Researchers are exploring new methods to address the limitations of existing models, such as CLIP, including the development of novel distillation techniques, patch generation-to-selection approaches, and global-local object alignment learning. These advancements have the potential to improve the performance of vision-language models in various tasks, including zero-shot classification, retrieval, and image generation. Notable papers in this area include: Seeing What Matters: Empowering CLIP with Patch Generation-to-Selection, which presents a concise yet effective approach to enhance CLIP's training efficiency while preserving critical semantic content. GenHancer: Imperfect Generative Models are Secretly Strong Vision-Centric Enhancers, which proposes a method to effectively extract fine-grained knowledge from generative models and mitigate irrelevant information. Zero-Shot Visual Concept Blending Without Text Guidance, which introduces a novel technique for fine-grained control over feature transfer from multiple reference images to a source image.