Conjugated Semantic Pool Improves OOD Detection with Pre-trained Vision-Language Models | Read Paper on Bytez