FARABI at WMT 2026: Post-Hoc Channel Controllers for Low-Resource Arabic–Asian Machine Translation

We describe our submission to the WMT26 Low-Resource Arabic-Asian Language Translation shared task covering ten translation directions between Arabic and five Asian languages. Our central question is whether a post-hoc controller can improve translation quality beyond a fine-tuned backbone without modifying any model weight. For the five into-Arabic directions, our primary system jointly fine-tunes facebook/nllb-200-3.3B on all source languages and augments it at inference time with per-direction NTK-Mirror controllers (fewer than 6K parameters per direction) that rescale frozen transformer output channels. For the five from-Arabic directions, the backbone is QWEN2.5-7B-INSTRUCT adapted via QLoRA, likewise augmented with NTK-Mirror. In the official evaluation, our system ranked first on Bengali- and Indonesian-to-Arabic, second on Hindi- and Urdu-to-Arabic, and third on English-to-Arabic, placing in the top three on all five into-Arabic directions. Post-submission analysis indicates that the weaker from-Arabic results stem primarily from a model-selection decision at submission time.