一作ACL 2026Poster· Natural Language Processing
From AR to Diffusion: Efficiently Adapting Large Language Models with Strictly Causal and Elastic Horizons
Xiangyu Ma · Teng Xiao · Zuchao Li · Lefei Zhang
摘要
Diffusion models promise efficient parallel text generation but rely on bidirectional attention, creating a structural mismatch with pre-trained Autoregressive (AR) models. This incompatibility precludes reusing robust AR priors, necessitating prohibitive pre-training from scratch. To bridge this...