Confidence Decoding

A decoding heuristic for masked diffusion language models in which, at each step, the position with the highest prediction confidence is always selected for revealing. It serves as a stable reference trajectory but disregards information gain.