SelFusion: Self-distillation for Diffusion Language Models
Experimental results on instruction-following tasks show that the proposed self-distillation framework, SelFusion, substantially outperforms other KD methods with external LLM and DLM teachers, providing a practical path toward improving DLM generation quality.
Hyeongsoo Lim, Jinyoung Kim, Eunjoo Seo et al.
· Annual Meeting of the Associ... · 0 citations