The field of vision-language models is rapidly evolving, with a focus on improving few-shot learning, zero-shot generalization, and fine-grained understanding. Researchers are exploring innovative approaches to adapt large pre-trained models to new tasks and datasets, including topology-aware tuning, testing-time distribution alignment, and compositional image-text matching. These methods aim to leverage the strengths of vision-language models while addressing their limitations, such as the inability to perform entity grounding and compositional matching. Notably, the use of multimodal models, large datasets, and meticulous training methods is becoming increasingly important for achieving state-of-the-art results. Overall, the field is moving towards more effective and efficient models that can capture subtle semantic differences and improve overall performance. Noteworthy papers include: TeDA, which proposes a novel framework for testing-time distribution alignment to adapt pre-trained 2D vision-language models for unknown 3D object retrieval. FG-CLIP, which enhances fine-grained understanding through large multimodal models, high-quality datasets, and carefully designed training methods. Compositional Image-Text Matching and Retrieval by Grounding Entities, which proposes a novel learning-free zero-shot augmentation of CLIP embeddings with favorable compositional properties.