Multimodal Generative AI in Education: A Systematic Literature Review of Tasks, Modalities, and Methods

Abstract

Background: Multimodal Generative Artificial Intelligence (AI) is transforming educational technology by enabling systems to generate and interpret multiple forms of information, including text, images, diagrams, audio, and video, primarily through multimodal large language models (MLLMs). Prior reviews of AI in education have focused on text-based large language models or a single instructional domain, without quantifying how multimodal systems are distributed across educational tasks, modalities, and methods. Objective: This systematic literature review addresses that gap by examining current applications of multimodal generative AI in education through three questions: (1) What educational tasks are targeted, and how are they distributed? (2) Which modality configurations (text, image, speech/audio, video) are used, and which remain underexplored? (3) What methodological approaches are employed, and how are they distributed? Methods: Thirty-eight papers published between 2020 and January 19, 2026 were reviewed using a two-level coding strategy consisting of dominant categories and secondary tags across educational tasks, modality configurations, and methodological approaches. Results: The literature primarily focuses on learner-facing feedback, tutoring, and learning guidance (17 of 38 studies, 44.7%), followed by assessment, scoring, and evaluation (10 of 38, 26.3%) and learning-support utilities (8 of 38, 21.1%). Most systems employ vision-language (text and image) configurations (31 of 38, 81.6%), while video-inclusive (4 of 38, 10.5%) and speech/audio-dominant configurations (3 of 38, 7.9%) remain comparatively infrequent. Most studies report technical performance metrics rather than validated pedagogical outcomes. Conclusion: Current educational applications of multimodal generative AI concentrate on vision-language tutoring and evaluation, while video and speech/audio modalities and pedagogically validated evidence remain comparatively underexplored. This review maps where educational MLLM work is concentrated and identifies speech/audio-rich interaction and video-grounded classroom understanding as priority directions for future multimodal educational systems.