TRAC: Length-Aware Decoding for Diffusion Machine Translation
November 20, 2026
TRAC is a training-free adaptive decoding method for diffusion language models that improves machine translation efficiency. It adjusts target generation length based on source length and translation direction to reduce redundant computation during denoising.
HOW THIS AFFECTS YOU
●
builderYou can get faster translation inference using diffusion models by reducing unnecessary token generation.
●
researcherThis introduces a way to adapt target lengths in diffusion-based text generation.