The field of multimodal large language models is rapidly advancing, with a focus on improving performance in tasks such as video question answering, open-world egocentric activity recognition, and referring video segmentation. Researchers are exploring various techniques, including reinforcement fine-tuning, probabilistic jump diffusion, and gaze consensus adaptation, to enhance the models' ability to reason and understand complex visual and temporal information. Notable papers include CrowdVLM-R1, which proposes a novel framework for crowd counting using fuzzy group relative policy reward, and ProbRes, which introduces a probabilistic residual search framework for open-world egocentric activity recognition. Other significant contributions include the development of InstructionBench, a benchmark for instructional video understanding, and RAVEN, an agentic framework for multimodal entity discovery from large-scale video collections. These advancements have the potential to significantly improve the performance of multimodal large language models in various applications, including video question answering, activity recognition, and content retrieval.