Search papers, labs, and topics across Lattice.
1
0
Reward hacking is mitigated and sample quality preserved in few-step discrete diffusion models by replacing intractable KL penalties with a novel inverse-distillation regularization.
