This paper describes NCI-MT’s submission to the low-resource Arabic-to-Asian language translation task as part of the WMT shared task. Our submission covered six of the ten language pairs. We built two separate systems, the primary and the contrastive. The primary was built using the data provided by WMT. In this approach, we fine-tuned our baseline model (NLLB) with per-direction LoRA adapters. For the contrastive system, we introduced an additional quality-gated refinement stage, which includes two stages: in the first stage, we use a reference-free quality estimator to evaluate the quality of the translations produced by our base model. In the second stage, we iteratively re-translate those translations that are below a certain threshold using LLM-based in-context learning. Our results show that both systems are successful in improving the translation quality in low-resource scenarios.

This work presents the details and findings of the first mentorship in speech translation (SpeechT), which took place in December 2024 and January 2025. To fulfil the mentorship requirements, the participants engaged in key activities, including data preparation, modelling, and advanced research. The participants explored data augmentation techniques and compared end-to-end and cascaded speech translation systems. The projects covered various languages other than English, including Arabic, Bengali, Galician, Indonesian, Japanese, and Spanish.