Verified, then logged.
Sun 22 Nov 2026 · 00:21 KST / 15:21 UTC한국어 | ENSearch /

Breaking01 / AI

TRAC speeds up diffusion translation by adapting text length

In the researchers’ tests, TRAC reduced model calls and lowered the macro-average COMET translation quality score by 0.02 points.

Seen from behind in an AI research lab, an anonymous researcher compares repetitive and context-dependent language-model responses on two monitors with unreadable screen content. AI-generated image

Researchers introduced TRAC on 20 Nov 2026, reporting that it generated sentences at up to 2.19 times the baseline speed when translating between Chinese and English in either direction.

According to the arXiv paper’s abstract, TRAC changes how diffusion models generate translations without additional training. It adjusts the output length based on the length of the source text and the direction of translation.

The researchers said TRAC finalises tokens, the units of text a model generates, early if their predictions remain stable across several steps, even when confidence is moderate.

In the researchers’ published tests, TRAC generated sentences at 1.91 to 2.19 times the speed of Fast-dLLM and made 24.9% to 36.8% fewer model calls. The COMET translation quality score fell by 0.02 points when averaged with equal weight across evaluation categories.

The abstract did not disclose performance figures for any other language pair.