Guiding Cross-Modal Representations with MLLM Priors via Preference Alignment | Read Paper on Bytez