The recent advancements in multimodal large language models (MLLMs) have been significant, with a notable focus on enhancing their capabilities in specialized domains such as electrical and electronics engineering, finance, and scientific research. These models are being increasingly tested and refined for their ability to handle complex, real-world scenarios that require a deep integration of visual and textual information. The field is moving towards developing benchmarks that not only assess the general capabilities of these models but also evaluate their performance in domain-specific tasks. This shift is crucial for advancing the practical applications of MLLMs in fields where visual complexity and specialized knowledge are paramount. Notably, the introduction of benchmarks like EEE-Bench and MME-Finance highlights the need for models that can understand and reason about intricate images and professional instructions, which are essential for tasks in engineering and finance. Additionally, the development of M3SciQA underscores the importance of multi-modal, multi-document scientific question answering, reflecting the complexity of real research workflows. These innovations are pushing the boundaries of what MLLMs can achieve, but also reveal critical limitations, such as the 'laziness' phenomenon observed in EEE-Bench, where models tend to rely on text over visual context. Overall, the field is progressing towards more robust and domain-specific evaluations, aiming to create models that can effectively handle the multifaceted demands of specialized fields.