The field of multimodal video understanding is rapidly evolving, with a focus on developing more sophisticated models that can effectively integrate visual and language information. Recent research has highlighted the importance of temporal understanding, with studies showing that traditional positional encodings may not be as crucial as previously thought, and that causal information pathways can emerge through inter-frame attention. Additionally, there is a growing interest in online video grounding, with models being designed to handle hybrid-modal queries and localize specific moments in videos. Another area of research is focused on improving the safety and reliability of video large language models, with studies identifying critical vulnerabilities in current designs and proposing new strategies for sampling and decoding. Noteworthy papers in this area include: Failures to Surface Harmful Contents in Video Large Language Models, which highlights the need for more robust sampling and decoding mechanisms to guarantee semantic coverage. Causality Matters: How Temporal Information Emerges in Video Language Models, which proposes efficiency-oriented strategies for staged cross-modal attention and temporal exit mechanisms. When and What: Diffusion-Grounded VideoLLM with Entity Aware Segmentation for Long Video Understanding, which introduces a diffusion temporal latent encoder and object grounded representations to enhance temporal perception and language-vision alignment.