Efficient Diffusion LLMs via Temporal-Spatial Parallel Decoding and Confidence Extrapolation

In three linesNew method to accelerate diffusion-based language models (dLLMs). Temporal-Spatial Parallel Decoding (TSPD) and Confidence Extrapolation (CE) reduce unnecessary denoising iterations by analyzing token-wise trajectories and forecasting future logit trends without additional training.

Summary generated by Claude — human-verified