Unlocking Multimodal Models with Lightweight Fine-Tuning

The recent proliferation of large-scale multimodal foundation models, which integrate vision and language understanding within a unified framework, has profoundly transformed the landscape of artificial intelligence, enabling unprecedented capabilities across tasks such as image captioning, visual question answering, cross-modal retrieval, and multimodal reasoning. Despite their impressive performance, these models present significant challenges for practical adaptation, as their massive parameter counts-often in the billions-render full fine-tuning computationally prohibitive, memoryintensive, and inefficient for deployment in resource-constrained environments. Parameter-efficient fine-tuning (PEFT) has emerged as a compelling solution to these challenges, offering a spectrum of strategies that introduce a relatively small number of task-specific parameters while keeping the majority of the pre-trained model frozen. This survey provides a comprehensive review of PEFT techniques applied to multimodal foundation models, focusing on three primary families: adapter-based methods, which insert lightweight bottleneck modules into hidden layers; low-rank adaptation methods, which efficiently modify linear projections within transformers; and prompting-based approaches, which manipulate input representations or latent activations to guide downstream behavior. We examine the mathematical underpinnings of these approaches, including formal definitions, parameter budgets, and the design of modality-specific adaptations, and highlight their empirical effectiveness across a wide range of multimodal tasks. In addition, we discuss critical challenges in PEFT, including optimal allocation of adaptation capacity across layers and modalities, robustness and generalization under distributional shifts, scalability in extremely large models, and theoretical understanding of task-specific low-dimensional manifolds within high-dimensional parameter spaces. Finally, we outline future research directions, emphasizing dynamic and adaptive fine-tuning mechanisms, cross-modal transferability, integration with efficiency-enhancing techniques such as pruning and quantization, and the pursuit of interpretable and theoretically grounded adaptation strategies. Through this survey, we aim to provide both a rigorous conceptual framework and practical guidance for leveraging PEFT to deploy flexible, efficient, and high-performing multimodal foundation models, bridging the gap between large-scale pre-training and real-world task adaptation while laying the groundwork for continued innovation in scalable, modular, and interpretable multimodal AI.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.