Federated prompt learning is a technique for adapting large vision-language models such as CLIP to multiple distributed clients. Instead of fine-tuning the entire model or exchanging raw data, each client works with lightweight soft prompts that are updated collaboratively. The arXiv abstract introduces FedRAPP as a new method in this space, focusing on test-time prompt tuning and reliability-aware prompt personalization.

The paper positions FedRAPP against existing federated prompt learning methods, which the abstract says have shortcomings. However, the available excerpt ends mid-sentence at "existing federated prompt learning methods are…", so the specific limitations are not stated in the source. Based on the title and the abstract's framing, the contribution appears to be a reliability-aware mechanism for personalizing prompts at test time, though full details are not available in the provided text.