Data Party THU
Nov 10, 2026 · Artificial Intelligence
One-Step Text Generation with Seq-Drifting: Diffusion LMs Break Multi-Step Barrier
Researchers from City University of Hong Kong, Shenzhen University, USTC, and Oxford propose Seq-Drifting, a context-dependent drifting field method enabling one-step text generation for diffusion language models, achieving 11.2x speedup over 4-step baselines with Gen-PPL of 36.01 on 128-token generation, and validated on translation, summarization, and reasoning tasks.
Seq-Driftingcontext-dependent drifting fielddiffusion language models
0 likes · 11 min read