The field of video understanding is moving towards more precise spatio-temporal reasoning, with a focus on incorporating physical information and multi-object layouts. This is being achieved through the development of novel graph-based methods and the introduction of new benchmarks and datasets that support the training of more advanced models. One notable trend is the growing importance of multimodal large language models (MLLMs) and their ability to capture complex spatial and temporal relationships. Another key area of research is the development of more robust and accurate models for procedure step recognition, which is critical for applications such as embodied intelligence and human-AI interaction. Noteworthy papers include: Video-STR, which presents a novel graph-based reinforcement method for precise video spatio-temporal reasoning, achieving state-of-the-art results on various benchmarks. LSVOS 2025 Challenge Report, which introduces a new track for complex video object segmentation and highlights emerging trends in the field, such as the growing role of LLM/MLLM components and memory-aware propagation. Learning to Recognize Correctly Completed Procedure Steps in Egocentric Assembly Videos, which proposes a dual-stream framework for procedure step recognition that leverages both spatial and temporal features, reducing the average delay between actual and predicted assembly step completions. SVAG-Bench, which introduces a novel task of spatio-temporal video action grounding and provides a large-scale benchmark and a baseline framework for joint spatial and temporal grounding.