RLHF-V’s IT

• Paper: RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedback
• GitHub Link
• Publisher: Arxiv
• Author Affiliation: Tsinghua University
• Type
  • SFT
  • RLHF
• Multi-turn
  • ✔
  • ✖
• Input Modalities $\rightarrow$ Output Modalities
  (I: Image, V: Video, A: Audio, 3D: Point Cloud, T: Text, B: Bounding box, Tab: Table, Web: Web page)
  • I+T $\rightarrow$ T
• Source
  • Collected human preference
• Method
  • Manu.
• I/V/A Scale
  • I
    • Not report
  • V
    • Not report
  • A
    • Not report
• Dialog Turn
  • Not report
• Instance Scale
  • 1.4K