The Dynamic Mixture of Curriculum LoRA Experts (D-MoLE) is a new method designed for continuous multimodal instruction fine-tuning, proposed by the Interactive Content Security Team of Alibaba Group Security Department and Tsinghua University on June 13, 2025. It aims to enable the multimodal large language model (MLLM) to continuously adapt to new tasks while effectively retaining existing knowledge under a limited parameter budget. The related paper results are "Dynamic Mixture of Curriculum LoRA Experts for Continual Multimodal Instruction Tuning", the paper has been included in ICML 2025.

D-MoLE combines the concepts of LoRa and Mixture of Experts (MoE) and introduces a curriculum learning mechanism. It dynamically selects and combines different LoRa modules to adapt to new tasks while minimizing interference with existing knowledge. Extensive experiments demonstrate that D-MoLE significantly outperforms state-of-the-art baseline models, achieving an average improvement of 15% over the best baseline. This is the first study of continuous learning in MLLMs from an architectural perspective.