Entropy-Valley Length-Adaptive Decoding for Masked Diffusion Translation
August 25, 2026
The Entropy-Valley (EV) method provides a training-free length selector for masked diffusion machine translation by scoring candidate target canvases via mean predictive entropy. EV recovers up to 65.3% of COMET-22 gains from reference lengths on En-Zh and Zh-En tasks compared to corpus statistics baselines.
HOW THIS AFFECTS YOU
●
builderYou can improve diffusion-based translation quality without retraining by implementing entropy-based length selection.
●
researcherThis demonstrates that denoising-friendly target lengths can differ from reference text lengths.