Search papers, labs, and topics across Lattice.
Tsinghua University
1
0
3
Multimodal models forget how to see and reason after SFT, but PRISM realigns them before RL, boosting performance by up to 6%.
