Entropy-Valley: Length-Adaptive Decoding for Diffusion MT
August 22, 2026
Entropy-Valley (EV) is a training-free method for selecting target canvas lengths in masked diffusion machine translation. By scoring candidate lengths using mean predictive entropy, EV recovers up to 65.3% of the COMET-22 gain from reference target lengths.
HOW THIS AFFECTS YOU
●
builderThis offers a way to improve the coverage and redundancy of diffusion-based language models without retraining.
●
researcherYou can improve diffusion-based translation by optimizing the target token length during the decoding process.