The field of video large language models (Video LLMs) is rapidly advancing, with a focus on improving the understanding of long videos. Researchers are exploring innovative approaches to address the challenges of processing large volumes of data and temporal complexity. One key direction is the development of adaptive frame selection and multi-resolution scaling methods, which enable Video LLMs to capture query-related crucial spatiotemporal clues effectively. Another area of research is the design of efficient video language models that can process extremely long videos in real-time, using techniques such as flash memory modules and moment sampling. These advancements have the potential to significantly improve the performance of Video LLMs on various video understanding tasks. Noteworthy papers in this area include Q-Frame, which introduces a novel approach for adaptive frame selection and multi-resolution scaling, and Flash-VStream, which proposes an efficient video language model capable of processing extremely long videos in real-time. Additionally, Temporal Chain of Thought presents an inference strategy that curates the model's input context to improve long-video understanding, and AuroraLong demonstrates the potential of efficient, linear RNNs to democratize long video understanding.