Adaptive Multimodal Intelligence under Operational Constraints: Models, Methods, and Open Challenges

In high-stakes areas, the decision-making process that takes place in real life has become more dependent on the multimodal foundation models that can process different types of data together, including text, images, audio, video, sensor data, and structured records. The authors of the present work are describing the progress in AI of multimodal kinds with a focus on the decision-making models under real-world constraints. They group the existing literature based on the fundamental technological characteristics like the fusion techniques of multiple modalities, choosing the modalities according to the resources that are available, being resistant to alterations in the data distribution and attacks, and being able to explain inputs that are of different types. The paper not only displays the application fields medical diagnosis, robots, self-driving cars, factory automation, and monitoring but it also points out the practical deployment issues, operational bottlenecks, and known failures. One of the things that these areas have in common is the need for proper evaluation frameworks that will assess technologies not only according to their standard accuracy metrics but also based on their decision-aware performance, risk sensitivity, and reliability. The paper also discusses moral and legal issues, such as data provenance, fair use and fine-tuning practices, and privacy risks associated with cross-modal inference. Moreover, it points out research problems that still exist in highly effective data curation, reasoning over a long time period, causation determination, lifecycle monitoring, and energy-saving modeling. The survey concludes with concrete recommendations aimed at guiding the responsible design, assessment, and implementation of multimodal decision-making systems.