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

Name: Continual Learning with Elastic Regularization and Synthetic Replay for Federated MLLM Fine-Tuning
Uploaded: 2026-10-06T07:36:35.447414+00:00
Duration: 267 s
Description: FedCMM addresses catastrophic forgetting in federated fine-tuning of Multimodal Large Language Models (MLLMs) by integrating three safeguards: modality-aware parameter regularization, data-level synthetic replay, and similarity-aware gradient aggregation. This holistic approach ensures robust adaptation to new tasks across distributed networks while preserving knowledge from previous tasks.

FedCMM addresses catastrophic forgetting in federated fine-tuning of Multimodal Large Language Models (MLLMs) by integrating three safeguards: modality-aware parameter regularization, data-level synthetic replay, and similarity-aware gradient aggregation. This holistic approach ensures robust adaptation to new tasks across distributed networks while preserving knowledge from previous tasks.

Open the full episode