DCE and SRCL proposed as an alternative to on-policy self-distillation
A post attributes the proposal to Meta's team and reports Qwen3-8B average accuracy rising from 30.76% to 65.97%.
TLDR
A post says Meta's team proposes DCE and SRCL as an alternative to on-policy self-distillation (OPSD). It reports that Qwen3-8B's average accuracy rose from 30.76% to 65.97%.
Combined views
47.9K
3 Sources, first seen 8h ago
496 likes
