SDPO (Self-Distillation)

SDPO ("Reinforcement Learning via Self-Distillation") turns rich textual feedback into a dense self-distillation training signal for RLVR, rather than relying only on sparse scalar rewards. It matches state-of-the-art reasoning performance with roughly 4× fewer generations, making verifiable-reward post-training substantially more sample-efficient. From Andreas Krause's LAS group (Hübotter/Guestrin/Krause); rapid adoption (~1K★ within months).