Continual Learning with Elastic Regularization and Synthetic Replay for Federated MLLM Fine-Tuning

Federated fine-tuning of Multimodal Large Language Models faces catastrophic forgetting, where sequential updates erase visual and linguistic knowledge. This is critical for safety-sensitive domains like content moderation. Meanwhile, a common AI misconception involves using fine-tuning for document knowledge. In reality, RAG acts as a library provided at query time for knowledge retrieval, whereas fine-tuning is intended for teaching a model specific writing styles or behaviors, preventing wasted GPU budgets on incorrect implementation paths.