The field of video-language understanding and generation is rapidly advancing, with a focus on improving temporal reasoning, multimodal understanding, and efficient computation. Researchers are exploring new architectures and training paradigms to enable more accurate and efficient video-language models. Notably, the use of hierarchical dual-stream architectures, temporal-aware fuseformers, and spatial-temporal rotary positional embeddings are showing promising results. Additionally, techniques such as layer caching, online cluster distillation, and evolutionary caching are being developed to accelerate inference and reduce computational costs. These advancements have the potential to enable more effective and efficient video-language understanding and generation systems. Noteworthy papers include: DaMO, which introduces a data-efficient Video LLM for accurate temporal reasoning and multimodal understanding. DISCOVR, which presents a self-supervised dual branch framework for cardiac ultrasound video representation learning. ReFrame, which explores layer caching for accelerated inference in real-time rendering. Deja Vu, which accelerates ViT-based VideoLMs by reusing computations across consecutive frames. VideoMAR, which proposes a concise and efficient decoder-only autoregressive image-to-video model with continuous tokens. EVA02-AT, which introduces spatial-temporal rotary positional embeddings and symmetric optimization for egocentric video-language understanding. Show-o2, which presents improved native unified multimodal models that leverage autoregressive modeling and flow matching. ECAD, which proposes evolutionary caching to accelerate diffusion models.