The field of multimodal research is moving towards more practical and real-world applications, with a focus on developing datasets and models that can handle complex scenarios. Researchers are working on creating evaluation metrics and benchmarks that can accurately assess the performance of multimodal models, particularly in areas such as concept customization and safety scenarios. Another area of focus is on improving the accuracy and reliability of multimodal models, including the development of new methods for calibrating judge models and handling multilingual translation tasks. Notable papers in this area include: PRIM: Towards Practical In--Image Multilingual Machine Translation, which proposes a new dataset and model for handling real-world in-image machine translation tasks. Calibrating MLLM-as-a-judge via Multimodal Bayesian Prompt Ensembles, which introduces a new method for calibrating judge models to improve their accuracy and reliability. Human Preference-Aligned Concept Customization Benchmark via Decomposed Evaluation, which proposes a novel evaluation method and benchmark dataset for assessing concept customization models.