RLAIF-V

Framework for aligning multimodal LLMs through open-source AI feedback for super GPT-4V trustworthiness. Core alignment technique used in MiniCPM-o. Accepted as CVPR 2025 Highlight.