Data Party THU

Oct 10, 2025 · Artificial Intelligence

How DPad Cuts Inference Time 61× While Boosting Accuracy in Diffusion LLMs

The article analyzes a recent Duke University paper that reveals a "scratchpad" mechanism in diffusion large language models, proposes the DPad method to prune redundant suffix tokens before decoding, and demonstrates up to 61.4× faster inference with unchanged or even improved accuracy across multiple benchmarks.