The arXiv abstract introduces Lumen, a translation system submitted to the WMT 2026 General Machine Translation Shared Task. The paper emphasizes instruction-specialized translation, where the model must follow user instructions to generate accurate translations rather than simply converting text.
To achieve this, the authors describe a training approach built on quality-aware training and on-policy distillation. These techniques are aimed at improving the model's ability to handle specialized translation requests. The abstract also states that the open-source model is available, though the provided text cuts off before listing the exact release location.
Because the abstract is truncated, details on evaluation results, baselines, and specific architectural choices are not included in the source. The paper's contribution, as described, is the combination of these training methods for the shared task, with an open-source release for further use.