We present our submission to the WMT 2026 Shared Task on Low-Resource Indic Language Translation, covering four language pairs (English-Assamese, English-Mizo, English-Khasi, and English-Tagin in both directions) spanning a wide range of resource levels, from moderately-resourced Assamese down to Tagin, which barely appears in existing multilingual MT systems at all. We compare three models (NLLB-200-1.3B, Sarvam-Translate, and TranslateGemma) across three setups: fine-tuning on the official shared-task data alone, fine-tuning with added external parallel data, and a reinforcement-learning stage (RLOO) that optimizes a combined BLEU/chrF++ reward on top of the fine-tuned model. Adding external parallel data helps most for Khasi, improving English-to-Khasi by 18 BLEU, and for Mizo, where our system places first among contrastive submissions in both directions; the same augmentation leaves Assamese essentially unchanged despite contributing twice as many sentence pairs, suggesting that data provenance matters more than volume at these scales. The RLOO system degrades performance on Assamese and Mizo, where supervised fine-tuning had already produced competitive systems, but improves over the primary system on both Khasi directions and on every metric for English-to-Tagin, where it obtains the highest METEOR score of any contrastive submission. Reward-based training therefore appears most useful precisely where supervised fine-tuning has least to work with.

Discourse-level translation remains a major challenge for machine translation (MT) systems, particularly for translation from Indian languages to English. This difficulty arises due to factors such as rich morphology, diverse syntactic structures, unmarked gender distinctions in pronouns, and the limited availability of discourse-aware training data. Existing evaluation benchmarks primarily focus on sentence-level translation quality and fail to capture important discourse phenomena such as pronoun resolution and lexical cohesion. To address this gap, we introduce IndicDISCO-MT, a parallel benchmark dataset covering translations from eight Indian languages such as Bengali, Gujarati, Hindi, Marathi, Kannada, Tamil, Telugu and Urdu to English. On top of this dataset, we are first to introduce DiscoAlign, a human annotated word to word alignment benchmark dataset, that captures correspondences between source and target words across languages. In addition, we propose two evaluation benchmarks, ProAlign and LexiAlign, designed to specifically assess the ability of large language models (LLMs) and MT systems to handle personal pronouns and lexical cohesion. Our evaluation of recent LLMs and MT systems on these benchmarks shows that although models achieve high overall translation quality, they still struggle to accurately preserve discourse-level phenomena. The proposed benchmarks provide a systematic framework for evaluating discourse aware translation and can facilitate the development of MT systems that generate more coherent and contextually consistent translations.