Length-Adaptive Decoding for Masked Diffusion Machine Translation
Researchers have proposed a new method for improving machine translation called Entropy-Valley (EV). EV is a training-free length selector that determines the optimal target length for masked diffusion language models. This is done by scoring candidate target canvases based on mean predictive entropy from all-mask forward passes and selecting the canvas the model is most prepared to fill. The authors claim that EV recovers 64.9%, 65.3%, and 33.0% of the COMET-22 gain from ref
Researchers have proposed a new method for improving machine translation called Entropy-Valley (EV). EV is a training-free length selector that determines the optimal target length for masked diffusion language models. This is done by scoring candidate target canvases based on mean predictive entropy from all-mask forward passes and selecting the canvas the model is most prepared to fill. The authors claim that EV recovers 64.9%, 65.3%, and 33.0% of the COMET-22 gain from reference target lengths on En$ o$Zh, Zh$ o$En, and En$ o$De respectively. However, their diagnostics show that denoising-friendly lengths do not necessarily match reference lengths.
---
Why it matters: This matters to AI researchers because it addresses a long-standing issue in masked diffusion machine translation: determining the optimal target length for faithful rendering of source tokens. The proposed Entropy-Valley method provides a practical solution to this problem, which can improve the performance of machine translation systems.
Source: https://arxiv.org/abs/2608.22274
This article was originally published at: https://arxiv.org/abs/2608.22274