One-to-many machine translation is traditionally computationally expensive for autoregressive systems due to linear latency scaling with sequence length and target language count. A new approach uses a discrete diffusion framework to refine all target languages simultaneously, achieving sublinear latency scaling relative to the number of targets.
Architecture and Training
The system operates as a single unified model conditioned on a continuous semantic anchor rather than source tokens. This design allows zero-shot transfer to unseen source languages without retraining, maintaining approximately 75% of supervised translation quality on these sources.
Performance Results
With accelerated sampling, the framework achieves comparable supervised quality to autoregressive baselines. It demonstrates a 2x speedup and an 11.9% improvement in zero-shot BLEU scores compared to standard autoregressive methods.
Source: https://arxiv.org/abs/2609.16312
