Prism: Bidirectional Self-Teaching for Unified Multimodal Models
Prism, a bidirectional self-teaching post-training method for UMMs that derives bidirectional supervision from self-generated interleaved rollouts, constructing training targets according to the reliability of their text and image chains according to their reliability states.
Peng Zhang, Qian Chen, Xian-Yin Zhang et al.
· 0 citations