Federated Multi-Label Prompt Tuning for Vision-Language Models

A new paper on arXiv (2605.28347) introduces FedMPT, the first method specifically designed for federated multi-label recognition (MLR) using vision-language models (VLMs). The approach addresses the problem of overfitting to spurious label correlations when VLMs are adapted to private, heterogeneous client data in decentralized settings. By applying a causal model with front-door adjustment, FedMPT decouples the MLR process through intermediate variables that amplify oracle label co-occurrence, steering the model toward generalizable conditions to reduce erroneous label activation.