The Mathematical Foundation and Optimization Methods of the Core Modules of Multimodal Large Models
Multimodal large models integrate language, vision and other input modalities through attention, alignment, connector mapping and optimization modules, and their capabilities depend not only on parameter scale but also on how these components exchange and transform information. This review examines the mathematical foundations of scaled dot-product attention, contrastive learning, positional encoding, low-rank adaptation, quantization and knowledge distillation. Attention scaling is explained through variance control, InfoNCE through mutual-information estimation, and parameter-efficient adaptation through low-rank approximation. Efficient attention computation and activation-aware quantization are also analysed from computational and numerical perspectives. Together, these connections clarify how architecture and computation shape stability, efficiency and reliability, and they identify directions for more interpretable and controllable multimodal systems.
Authors
- Zhuoxuan Cai (ORCID: https://orcid.org/0009-0001-6930-4016)
Institutions
- Northeastern University (CN)
Publication Details
- Journal
- Theoretical and Natural Science
- Published
- 2026-11-17
- DOI
- https://doi.org/10.54254/2753-8818/2027.ta39198
- Primary Topic
- Multimodal Machine Learning Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00