FedDTL decouples VLM encoders across server and clients with modality alignment and uses two-stage local fine-tuning (supervised then RL) to balance global adaptation and generalization in heterogeneous federated learning.
However, the accuracy on novel classes with mismatched backbones (i.e., ViT-L/14 and ViT-B/32) declines significantly, demonstrating generalization degradation
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Decoupled Training with Local Reinforcement Fine-Tuning in Federated Learning
FedDTL decouples VLM encoders across server and clients with modality alignment and uses two-stage local fine-tuning (supervised then RL) to balance global adaptation and generalization in heterogeneous federated learning.