Researchers present Entropy-Valley, a training-free technique for selecting target canvas lengths in masked diffusion machine translation that scores candidate lengths using mean predictive entropy from all-mask forward passes to pick the canvas the model is best prepared to fill. Evaluated on English-Chinese and English-German translation, the method recovers 33-65% of the performance gains achievable with reference target lengths and matches or exceeds a comparable autoregressive baseline. The results indicate that deciding which tokens to reveal first matters less than how the target length is supplied in masked diffusion translation.
