RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-Grained Correctional Human Feedback | IEEE Conference Publication | IEEE Xplore