Aashish Dhawan, Christopher Driggers-Ellis, Dzmitry Kasinets, Christan Grant, Daisy Zhe Wang
WMT 2026 · EMNLP, Budapest · 2026
#1 in 10 directions and #2 in 8 more across the WMT26 primary leaderboards at EMNLP 2026, Budapest
This paper describes the University of Florida Gators submission to the WMT26 Low-Resource Indic Language Translation shared task. We adapt the retrieval-augmented many-shot translation pipeline from our AmericasNLP 2026 system to translate between English and eleven North-Eastern Indian languages in both directions. At inference time, BM25 retrieves the most similar parallel examples from a language-specific training bank, and Gemini 2.5 Flash translates the input conditioned on these examples. No model fine-tuning is involved. Training banks combine official WMT26 data with publicly available corpora such as Samanantar and prior WMT shared task releases. A grid search over retrieval count r and development exemplar count d across all 22 language-direction pairs selects the best configuration for each submission.
The system adapts the retrieval-augmented translation recipe from our AmericasNLP 2026 work to a broader multilingual shared-task setting. Instead of fine-tuning a translation model, it retrieves the most relevant parallel examples for each input with BM25 and conditions Gemini 2.5 Flash on those examples in a many-shot prompt.
Because the method is retrieval-driven, each language direction gets its own training bank and its own prompt configuration. This keeps the pipeline lightweight while still letting it specialize to low-resource translation pairs through example selection.
Training banks are built by combining the official WMT26 shared-task data with public multilingual corpora including Samanantar and earlier WMT releases. That gives the retriever a larger pool of candidate demonstrations while staying grounded in task-relevant parallel data.
For submission tuning, the system runs a grid search over retrieval count and development exemplar count across all 22 language-direction pairs. The final configuration for each direction is selected from development performance rather than applying a single shared prompt budget everywhere.
The shared task is part of EMNLP 2026 in Budapest as part of the Eleventh Conference on Machine Translation (WMT26), scheduled for October 28-29, 2026. In the published results booklet, the University of Florida Gators primary system finishes #1 in 10 translation directions and #2 in 8 more, placing in the top two for 18 of the 20 reported directions.
That spread is more impressive than a single headline win: the system tops both English→Assamese and Assamese→English, leads several other directions including Mizo, Manipuri, Bodo, and Karbi settings, and stays consistently near the front almost everywhere else. It is a strong demonstration that well-designed retrieval-augmented inference can compete at the very top of a shared task in genuinely low-resource settings.
Back to Aashish Dhawan's site