Research & academia
Predict, Don't Iterate: Efficient Adaptive-Length Infilling for Diffusion Language Models
PILL targets a central infilling problem in diffusion language models: choosing the span length without a preset starting length or repeated, costly searches.
Why it matters: For infilling, selecting a starting length can affect both output quality and inference cost. PILL is presented as a way to avoid preset-length sensitivity while reducing extra forward passes; the abstract reports better benchmark results and faster running time than the strongest baseline.