TRAC cuts diffusion translation decoding steps with length-aware scheduling
A training-free decoding method for diffusion language models reduces forward calls in Chinese-English machine translation while keeping quality nearly unchanged, according to a new arXiv paper.
In plain words
Diffusion language models generate text by refining many positions at once, but they often compute more than needed. TRAC is a decoding adjustment that shortens the generated length and commits tokens earlier once their predictions stay stable across steps. The authors report faster translation with almost no quality loss.
What was announced
TRAC, described in a new arXiv paper, targets two sources of waste in diffusion machine translation. Fixed target generation lengths spend computation on unused positions, and confidence-only commitment holds back medium-confidence tokens whose predictions have largely stabilized. The method is training-free: it sets the target length from the source length and translation direction, falling back to the maximum length when outputs would be empty or unusually short, and it commits medium-confidence tokens early when cross-step top-1 persistence and approximate Jensen-Shannon divergence over Top-K distributions indicate stability.
On bidirectional Chinese-English translation across FLORES-200, WMT23, and NTREX-128, the authors report that TRAC reduces model forward calls by 24.9% to 36.8% and speeds up sentence-level generation by 1.91 to 2.19 times over Fast-dLLM, with a 0.02-point decrease in macro-average COMET. These figures come from the paper's abstract alone, so the underlying experimental setup, baselines beyond Fast-dLLM, and evaluation details cannot be verified from the supplied excerpt.
What it does not show
- Evidence is limited to the paper's abstract; experimental setup, datasets' roles, and full results are not available in the supplied passages.
- The reported speedups and quality figures are the authors' own measurements and have not been independently confirmed.
- The evaluation covers bidirectional Chinese-English translation only; results for other language pairs are not established.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.