Alignment in Multimodal Large Language Models: A Survey

Multimodal large language models (MLLMs) can generate plausible responses without using the relevant evidence or meeting task constraints. We separate the Modality Gap in cross-modal correspondence and evidence use from the Intent Gap in behavior relative to specified constraints, goals, and human expectations. Six dimensions organize the review: architecture, objective functions, data construction, training, evaluation, and applications. The first four control evidence access, optimization targets, supervision, and updates. Comparing mechanisms and supervision assumptions shows what interface compatibility, target imitation, preference learning, and verifiable rewards can each demonstrate about alignment. Evaluation at the representation, task, and behavior levels tests these choices through controlled input changes and reward-model validation, and we interpret each gain together with its model, task, and evaluation setup instead of comparing scores across conditions. Studies in robotics, safety, healthcare, recommendation, and climate decision support share three requirements: traceable evidence, valid supervision, and constraint-sensitive decisions. Stronger representations or higher benchmark scores can leave the claimed alignment mechanism unresolved, which motivates four research directions: evidence-consistent alignment across modalities, verified self-correction and test-time alignment, targeted repair from failure diagnosis, and reliable transfer to underrepresented settings.