Why treat all classes equally when picking prompt weights? Researchers from Univ. of Melbourne, Southeast Univ., & RIKEN present CARPRT. It scores each prompt’s relevance per class by averaging over predicted images—no training needed. Outperforms class-independent methods