作者位次 #3/5ACL 2026Poster· Natural Language Processing
S el F usion: Self-distillation for Diffusion Language Models
Hyeongsoo Lim · Jinyoung Kim · Eun Seo Seo · Min Ho Jang · Ji Won Yoon
摘要
Diffusion language models (DLMs) alleviate the inherent latency bottleneck of autoregressive (AR) large language models (LLMs), but their degraded generation quality limits practical applicability. Although knowledge distillation (KD) can be a promising direction for improving performance, we emp...