BM25-Augmented Many-Shot Translation for Low-Resource North-Eastern Indian Languages

# BM25-Augmented Many-Shot Translation for Low-Resource North-Eastern Indian Languages
Source: [https://arxiv.org/html/2608.13722](https://arxiv.org/html/2608.13722)
Christopher Driggers\-EllisUniversity of Floridadriggersellis\.cw@ufl\.eduDzmitry KasinetsUniversity of Floridadkasinets@ufl\.eduChristan GrantUniversity of Floridachristan@ufl\.eduDaisy Zhe WangUniversity of Floridadaisyw@cise\.ufl\.edu

###### Abstract

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[8](https://arxiv.org/html/2608.13722#bib.bib2);[3](https://arxiv.org/html/2608.13722#bib.bib25)to translate between English and eleven North\-Eastern Indian languages in both directions\. At inference time, BM25[38](https://arxiv.org/html/2608.13722#bib.bib14)retrieves the most similar parallel examples from a language\-specific training bank, and Gemini 2\.5 Flash[5](https://arxiv.org/html/2608.13722#bib.bib9)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[35](https://arxiv.org/html/2608.13722#bib.bib24)and prior WMT shared task releases\. A grid search over retrieval countrrand development exemplar countddacross all 22 language–direction pairs selects the best configuration for each submission\. Find our code on[Github](https://github.com/dhawan98/Gators_wmt26/tree/main)\.

11footnotetext:∗Equal contribution\.## 1Introduction

The WMT26 Low\-Resource Indic Language Translation shared task[28](https://arxiv.org/html/2608.13722#bib.bib8)covers eleven North\-Eastern Indian languages that span four language families: Assamese \(Indo\-Aryan\), Khasi \(Austroasiatic\), eight Tibeto\-Burman languages \(Mizo, Manipuri in both Bengali and Meitei Mayek scripts, Bodo, Kokborok, Karbi, Nyishi, and Tagin\), and Nagamese, an English\-based creole\. Several of these — particularly Nyishi[17](https://arxiv.org/html/2608.13722#bib.bib26), Karbi, and Tagin — have almost no prior MT work\.

The system we describe here emerges from two earlier projects\. In our LoResMT 2026 paper[7](https://arxiv.org/html/2608.13722#bib.bib1), we found that augmenting mBART with NLLB\-200\-generated synthetic data improved ChrF\+\+ for low\-resource Indigenous American languages, but the gains were modest and domain\-sensitive\. We then moved to a retrieval\-augmented approach for the AmericasNLP 2026 captioning task[8](https://arxiv.org/html/2608.13722#bib.bib2);[3](https://arxiv.org/html/2608.13722#bib.bib25): instead of fine\-tuning, we used BM25 to pull relevant parallel examples into a long Gemini 2\.5 Flash[5](https://arxiv.org/html/2608.13722#bib.bib9)prompt and let the model translate in context\. That system won the overall shared task competition across five Indigenous American languages, and the gap over fine\-tuned mBART was large enough that we wanted to test whether the approach would transfer to a different language family and task domain\.

For WMT26, the adaptation is straightforward\. The AmericasNLP pipeline had two stages — VLM image captioning in Spanish, then retrieval\-augmented Spanish\-to\-LoRes translation\. Since WMT input is already English text, we drop the vision stage and feed source sentences straight into the BM25\+Gemini pipeline\. The rest carries over: per\-language system prompts, ther,dr,dhyperparameter sweep, and the same scoring infrastructure\.

Figure 1:Overview of thegatorspipeline\.
## 2Background

The WMT Low\-Resource Indic shared task has run annually since 2023[29](https://arxiv.org/html/2608.13722#bib.bib5), growing from four language pairs to eleven in 2026[28](https://arxiv.org/html/2608.13722#bib.bib8)\. The task splits languages into two categories: Category 1 \(moderate data\) includes Assamese, Mizo, Khasi, Manipuri in Bengali script, Manipuri in Meitei Mayek, and Nyishi; Category 2 \(very limited data\) includes Bodo, Kokborok, Karbi, Nagamese, and Tagin\. The Indic shared task organizers produce no official ranking of submissions, nor is a single winner declared, but they do report numerous metrics including BLEU[31](https://arxiv.org/html/2608.13722#bib.bib17), ChrF\+\+[33](https://arxiv.org/html/2608.13722#bib.bib16), COMET[37](https://arxiv.org/html/2608.13722#bib.bib18), and others\.

Previous editions[27](https://arxiv.org/html/2608.13722#bib.bib6);[26](https://arxiv.org/html/2608.13722#bib.bib7)have been dominated by fine\-tuned multilingual models — IndicTrans2[13](https://arxiv.org/html/2608.13722#bib.bib12), mBART[22](https://arxiv.org/html/2608.13722#bib.bib13), and NLLB\-200[6](https://arxiv.org/html/2608.13722#bib.bib10)— combined with back\-translation[39](https://arxiv.org/html/2608.13722#bib.bib19)or transfer from related higher\-resource Indic languages \(see[36](https://arxiv.org/html/2608.13722#bib.bib20)for a survey\)\.

Our submission takes a different route\. The translation budget goes entirely to Gemini API calls, with BM25 handling example selection at inference time\. This follows work on retrieval\-augmented in\-context translation[43](https://arxiv.org/html/2608.13722#bib.bib22);[1](https://arxiv.org/html/2608.13722#bib.bib23)showing LLMs can match or exceed fine\-tuned baselines given well\-chosen parallel demonstrations[2](https://arxiv.org/html/2608.13722#bib.bib21)\. Related work from our lab on multimodal low\-resource data generation[45](https://arxiv.org/html/2608.13722#bib.bib4)and cross\-script parallel data[9](https://arxiv.org/html/2608.13722#bib.bib3)also informs the data preparation\.

## 3System Description

The overview of the pipeline is given in[Figure1](https://arxiv.org/html/2608.13722#S1.F1)\. In this section, we will describe the phases of this pipeline from start to finish\.

### 3\.1Data Preparation

The retrieval bankℛ\\mathcal\{R\}is built from WMT26 training data and augmented with WMT23 pairs and publicly available parallel corpora\. We retain 21 Manipuri and 3,554 Mizo pairs unique to WMT23[29](https://arxiv.org/html/2608.13722#bib.bib5)\.

#### Category 1 Languages

For Assamese, we add 142,080 pairs from Samanantar[35](https://arxiv.org/html/2608.13722#bib.bib24), 1,359 from Endangered Recipes Translated 500[19](https://arxiv.org/html/2608.13722#bib.bib28);[18](https://arxiv.org/html/2608.13722#bib.bib29);[16](https://arxiv.org/html/2608.13722#bib.bib30), and 2,009 from IndicGenBench FLORES[41](https://arxiv.org/html/2608.13722#bib.bib31)to the 54,000 WMT26 pairs\. Khasi receives 1,048 Endangered Recipes pairs in addition to 26,000 WMT26 pairs, while Manipuri receives 2,009 FLORES pairs and 21 WMT23 pairs in addition to 23,687 WMT26 pairs\.

We add no WMT\-external data for the remaining Category 1 languages\. Meitei\-Mayek has 17,001 WMT26 pairs\. Nyishi has 60,000 WMT26 pairs, and Mizo has 53,543 pairs across WMT26 and WMT23—49,989 and 3,554 pairs, respectively\. We focus our data\-collection efforts on the more resource\-constrained Category 2 languages, although additional parallel data may exist for these three languages\. Remaining paragraphs in this subsection will detail data augmentation for each Category 2 language separately\.

#### Bodo

We combine 15,214 WMT26 pairs with 134,931 pairs from the English–Bodo Parallel Dataset[30](https://arxiv.org/html/2608.13722#bib.bib32), 1,497 Endangered Recipes pairs, and 2,009 FLORES pairs to produce the largest corpus of WMT and external data\.

#### Karbi

We add 382 pairs mined fromA Grammar of Karbi[21](https://arxiv.org/html/2608.13722#bib.bib33)and 1,144 fromKarbi Texts[20](https://arxiv.org/html/2608.13722#bib.bib34)to the 1,000 WMT26 pairs\.

#### Kokborok, Nagamese, and Tagin

The corresponding WMT26 sets contain 2,269, 1,998, and 5,344 pairs\. We add 6,078 Kokborok pairs from the Kokborok Digitalisation Project[4](https://arxiv.org/html/2608.13722#bib.bib35);[15](https://arxiv.org/html/2608.13722#bib.bib36), 8,771 English–Nagamese pairs from FNPC\-25[23](https://arxiv.org/html/2608.13722#bib.bib37), and 854 English–Tagin pairs from GinLish v0\.1\.0[10](https://arxiv.org/html/2608.13722#bib.bib38)\.

All sources are merged using SHA\-256 source\-side deduplication, Unicode normalization \(NFC for Bengali and Devanagari and NFKC for Latin scripts\), and length\-ratio filtering\. For noisier external sources, we optionally run LaBSE[12](https://arxiv.org/html/2608.13722#bib.bib11)cross\-lingual similarity filtering\.

### 3\.2In\-Domain Data Filtering

We restrict retrieval to in\-domain data because sentence pairs from unrelated domains have little lexical or semantic overlap with the WMT26 inputs\. As a result, BM25 assigns them low relevance scores, making them unlikely to be selected as useful in\-context examples\. We use filtered, in\-domain data in both our primary and contrastive systems\. The final primary corpus sizesTrainand number ofPairs, including contrastive data, appear in[Table1](https://arxiv.org/html/2608.13722#S3.T1)\.

Cat\.LanguageISOTrainAug\.Pairs1AssameseAs52,4403\.3×\\times171,794MizoLus49,2931\.0×\\times50,620KhasiKha25,4950\.9×\\times22,847ManipuriMni22,5391\.0×\\times22,544MeiteiMni\_Mtei16,3121\.0×\\times16,312†\\daggerNyishiNjz49,6131\.0×\\times49,7932BodoBodo12,2689\.3×\\times114,410KokborokTrp2,0401\.0×\\times2,102KarbiMjw8641\.0×\\times899NagameseNag1,7985\.7×\\times10,169TaginTgj4,7030\.9×\\times4,421Table 1:Sizes of the in\-domain training data before and after augmentation\. The primary system uses theTraindata, while the contrastive system uses thePairsdata\. The augmentation ratio is computed as the number of augmented sentence pairs divided by the number of base pairs\.†\\daggerPairsrecalculated fromTrain\. They are equal because we use no external data for Mni\_Mtei\.
### 3\.3BM25 Retrieval

The source side ofℛ\\mathcal\{R\}is indexed with BM25 Okapi[38](https://arxiv.org/html/2608.13722#bib.bib14)\(k1=1\.5k\_\{1\}=1\.5,b=0\.75b=0\.75\)\. Given a source sentenceqq, the system retrieves the top\-rrtraining pairsℛr​\(q\)⊆ℛ\\mathcal\{R\}\_\{r\}\(q\)\\subseteq\\mathcal\{R\}, skipping exact matches on the query and deduplicating by 60\-character normalized prefix to prevent overlapping corpora \(e\.g\., Samanantar vs\. PMIndia\) from flooding the prompt with near\-identical examples\.

Getting the index direction right matters\. For En→\\toX, the index is built on the English sentences\. For X→\\toEn, the query is in the Indic language, so the index must be built on the Indic\-language side\. We had a bug in the AmericasNLP system[8](https://arxiv.org/html/2608.13722#bib.bib2)where Spanish \(Es\) was always indexed regardless of direction, but it was overlooked because X→\\toEs translation was not necessary\. We find and resolve the issue in this shared task\.

Configurations additionally includedddevelopment exemplars𝒟d⊆𝒟\\mathcal\{D\}\_\{d\}\\subseteq\\mathcal\{D\}, selected by word\-overlap ranking with leave\-one\-out exclusion\. The prompt context becomesP⁡\(q\)=𝒟d∪ℛr​\(q\)P\(q\)=\\mathcal\{D\}\_\{d\}\\cup\\mathcal\{R\}\_\{r\}\(q\)\. We conduct a grid search overr∈\{0,20,40,80\}r\\in\\\{0,20,40,80\\\}andd∈\{0,10,20,49\}d\\in\\\{0,10,20,49\\\}for each language and direction — a 4×\\times4 grid of 16 configurations per language–direction pair, evaluated on 50 held\-out development sentences per cell\. The grid is complete across all 22 language–direction pairs\.

### 3\.4Translation

Prompts go to Gemini 2\.5 Flash[5](https://arxiv.org/html/2608.13722#bib.bib9)via thegoogle\-genaiSDK \(temperature 0, thinking disabled, max output 512 tokens\)\. We use a 2\-key API pool with round\-robin and quota failover to manage rate limits\. Each of the eleven languages has its own system prompt encoding script, word order, and morphological type — the Khasi prompt warns about VOS order and diacritics, the Bodo prompt states that despite Devanagari script the language is Tibeto\-Burman and not Hindi\. We elaborate these prompting considerations in our appendix\.

Post\-processing strips model\-generated prefixes, applies Unicode normalization, and for non\-Latin scripts validates that output characters fall within the expected Unicode block\. The validation is especially important for Meitei Mayek \(U\+ABC0–U\+ABFF\), where Gemini occasionally falls back to Bengali\. A negligible fraction of output lines \(4 of 21,599; 0\.018%\) are empty due to Gemini safety filters on specific political sentences that temperature\-perturbed retry does not resolve\.

### 3\.5Relation to the AmericasNLP System

As shown in Figure[1](https://arxiv.org/html/2608.13722#S1.F1), the main architectural change from the AmericasNLP pipeline[8](https://arxiv.org/html/2608.13722#bib.bib2)is the removal of the vision\-captioning stage\. We also add bidirectional translation support through the BM25 index fix described, and merge multiple external corpora into retrieval banks\.

## 4Languages

Most language tracks in the shared task—eight of eleven—represent Tibeto\-Burman languages, while the three non\-Tibeto\-Burman languages are Assamese, Khasi, and Nagamese[34](https://arxiv.org/html/2608.13722#bib.bib42)\. The languages are predominantly Subject–Object–Verb \(SOV\), consistent with the clause\-final pattern common across the Tibeto\-Burman and local Indo\-Aryan languages of Northeast India\. Khasi, an Austroasiatic language, is the principal exception, with basic Subject–Verb–Object \(SVO\) order[34](https://arxiv.org/html/2608.13722#bib.bib42);[40](https://arxiv.org/html/2608.13722#bib.bib43)\. Assamese is an Indo\-Aryan language, while Nagamese is an Assamese\-lexified creole spoken across Nagaland[34](https://arxiv.org/html/2608.13722#bib.bib42);[24](https://arxiv.org/html/2608.13722#bib.bib41)\. Manipuri appears in two task tracks, one written in Bengali script and the other in Meitei Mayek[42](https://arxiv.org/html/2608.13722#bib.bib45)\. These tracks represent two scripts of the same language rather than genetically distinct languages; we refer to the Meitei Mayek track as Meitei elsewhere in the paper\. Bodo belongs to the Tibeto\-Burman family but is written in the Devanagari script[34](https://arxiv.org/html/2608.13722#bib.bib42);[25](https://arxiv.org/html/2608.13722#bib.bib44)\. In our experiments, Gemini occasionally produced Hindi\-like Bodo output, likely because Bodo shares the Devanagari script with Hindi and Nepali[25](https://arxiv.org/html/2608.13722#bib.bib44)\. Explicitly identifying Bodo in the system prompt removed this behavior\.

## 5Experimental Setup

Experiments run on the University of Florida HiPerGator cluster\. The translation pipeline is CPU\-only \(it makes API calls to Gemini\) and runs on thehpg\-defaultpartition\. The environment is a Python venv \(not conda\)\. GPU nodes \(hpg\-b200\) and mamba \(conda\) environments are used only for NLLB\-200 data generation Evaluation uses SacreBLEU to measure ChrF[32](https://arxiv.org/html/2608.13722#bib.bib15)and ChrF\+\+[33](https://arxiv.org/html/2608.13722#bib.bib16)\.

Since no official WMT26 development set was released early on, we hold the last 500 pairs or fewer from each language’sTraindata as the set of possible development exemplars𝒟\\mathcal\{D\}\(\|𝒟\|≤500\|\\mathcal\{D\}\|\\leq 500\) before adding external data to obtainPairs\. We record\|𝒟\|\|\\mathcal\{D\}\|for each language in[Table2](https://arxiv.org/html/2608.13722#S5.T2)\. The pilot grid search evaluates each\(r,d\)\(r,d\)cell onn=50n=50sentences from this split\. We submit the configuration that achieves the greatest ChrF score\.

Table 2:\|𝒟\|\|\\mathcal\{D\}\|for each language in the shared task\. Note that\|𝒟\|=500\|\\mathcal\{D\}\|=500except for some Cat\. 2 languagues\.
## 6Results

Invariantdev En→\\toXdev X→\\toEnCat\.LanguageISOAug\.PairsrrddChrFChrF\+\+rrddChrFChrF\+\+1AssameseAs3\.3×\\times171,794801052\.9549\.81404968\.8067\.98MizoLus1\.0×\\times50,620802053\.4051\.90402063\.5362\.53KhasiKha0\.9×\\times22,847804956\.4455\.71804975\.6475\.09ManipuriMni1\.0×\\times22,544804966\.7262\.34204969\.9767\.86MeiteiMni\_Mtei1\.0×\\times16,312801041\.2236\.86401055\.2052\.85NyishiNjz1\.0×\\times49,793802044\.9840\.58802058\.7657\.962BodoBodo9\.3×\\times114,410804976\.3772\.14201079\.6278\.38KokborokTrp1\.0×\\times2,102804954\.6951\.61804986\.9086\.26KarbiMjw1\.0×\\times899404949\.1944\.84404962\.7160\.82NagameseNag5\.7×\\times10,169802055\.7052\.59802066\.5964\.69TaginTgj0\.9×\\times4,421801046\.0243\.33402047\.0545\.77Table 3:Development contrastive submission results for each combination of language and translation direction\. Two optimalr,dr,dpairs are selected for each language from our search grid, one per translation direction\.Augmentation \(Aug\.\)gives a rough multiplier on the original size of the development set that our contrastive data augmentation provides\.Pairsis the total number of contrastive retrieval exemplars\|ℛ\|\|\\mathcal\{R\}\|after augmentation\.Invarianttest En→\\toXtest X→\\toEnCat\.LanguageISOAug\.PairsrrddBLEUChrF\+\+rrddBLEUChrF\+\+1AssameseAs3\.3×\\times171,794801020\.7752\.81404927\.2757\.35MizoLus1\.0×\\times50,620802021\.6050\.69402033\.2863\.36KhasiKha0\.9×\\times22,847804929\.1453\.16804925\.7652\.92ManipuriMni1\.0×\\times22,54480498\.1539\.89204945\.9372\.37MeiteiMni\_Mtei1\.0×\\times16,312801016\.6252\.85401038\.1366\.16NyishiNjz1\.0×\\times49,7938020––8020––2BodoBodo9\.3×\\times114,410804914\.6148\.21201034\.1062\.10KokborokTrp1\.0×\\times2,10280496\.0233\.70804916\.6741\.58KarbiMjw1\.0×\\times89940495\.7433\.854049––NagameseNag5\.7×\\times10,169802016\.6254\.14802019\.4844\.66TaginTgj0\.9×\\times4,42180103\.3124\.1340204\.0117\.73Table 4:Test contrastive submission results for each combination language and translation direction[28](https://arxiv.org/html/2608.13722#bib.bib8)\. Two optimalr,dr,dpairs are selected for each language from our search grid, one per translation direction\.Augmentation Ratio \(Aug\.\)gives a rough multiplier on the original size of the development set that our contrastive data augmentation provides\.Pairsis the total number of contrastive retrieval exemplars\|ℛ\|\|\\mathcal\{R\}\|after augmentation\. Entries inboldindicate that our submission performs best in the shared task for that metric\.In this section, we introduce the results of our method for translation of Category 1 and Category 2 languages in both the En→\\toX and X→\\toEn translation tasks\.[Table3](https://arxiv.org/html/2608.13722#S6.T3)lists the performance metrics for our contrastive approach for each language pair and translation direction, evaluated with the augmented development datasets prior to submission\. We measure performance using the ChrF and ChrF\+\+ metrics\. The former determined whatr,dr,dpairs feature in our submissions, and we include ChrF\+\+ for direct comparison of dev and test performance\.

[Table4](https://arxiv.org/html/2608.13722#S6.T4)lists contrastive performance for the languages and translation directions according to the BLEU and ChrF\+\+ measurements published by the shared task organizers[28](https://arxiv.org/html/2608.13722#bib.bib8)\. Other test\-time metrics are available in the shared task findings paper[28](https://arxiv.org/html/2608.13722#bib.bib8)\.

## 7Ablations

In this section, we look at alternative approaches to the proposed method\. We hope these ablations will guide future work by showing how to improve or not to improve on the current method and by deepening the discussion to follow about its strengths and weaknesses\.

### 7\.1Multi30k Domain Mismatch

The clearest negative result concerns synthetic data augmentation, which we apply to positive effect in multiple previous MT experiments[7](https://arxiv.org/html/2608.13722#bib.bib1);[8](https://arxiv.org/html/2608.13722#bib.bib2)\.

We prepare but ultimately*exclude*synthetic data that would have originally augmented three of the eleven languages in the present task\. Following the MultiScript30k methodology[9](https://arxiv.org/html/2608.13722#bib.bib3), we generate NLLB\-200 translations of the Multi30k image captions[6](https://arxiv.org/html/2608.13722#bib.bib10);[11](https://arxiv.org/html/2608.13722#bib.bib27)into Assamese, Mizo, and Manipuri \(∼\\sim29K pairs each\)\. This augmentation is the single largest driver of Guaraní performance in our AmericasNLP system\. Multi30k captions are somewhat domain\-matched to the captioning task[8](https://arxiv.org/html/2608.13722#bib.bib2)because they are image captions, and the queries at the MT stage of our submission are machine\-generated Spanish captions of images\. Meanwhile, WMT26 test data are sports news and government proceedings — “PCB requests ICC review,” “India loses 0\-8 in hockey”\.

We conclude that domain match contributes more strongly to performance than sheer volume in BM25\-based RAG and exclude MultiScript30k augmentation from all final retrieval banks\.

Figure 2:Dev\-to\-test performance changes in the contrastive system\.Left:En→\\toX translation performance for dev \(blue\) and test \(orange\) data\.Right:X→\\toEn translation performance for dev \(blue\) and test \(orange\) data\.Table 5:Results of repeating ourr,dr,dgrid search on near\-duplicate\-filtered WMT26 data\.Augmentation \(Aug\.\)gives a rough multiplier on the original size of the development set that our contrastive data augmentation provides\.Pairsis the total number of contrastive retrieval exemplars\|ℛ\|\|\\mathcal\{R\}\|after augmentation\. Unchanged optimalr,dr,dconfigurations initalics\.
### 7\.2Key Observations from the Grid Search

#### Script effect\.

While they are treated as entirely separate languages by our method and the shared task documentation, we have already acknowledged that Manipuri is represented twice in the present task\. What we call Manipuri is Manipuri in Bengali script while the language called Meitei is Manipuri written in the Meitei Mayek script\. For the Bengali script, our submission achieves 62\.34 dev ChrF\+\+ in the En→\\toMni direction, versus 36\.86 dev ChrF\+\+ for En→\\toMni\_Mtei translation\. This trend repeats for translation in the X→\\toEn direction but performance drops less steeply from 67\.86 to 52\.85 ChrF\+\+\. This represents roughly 25\- and 15\-point gaps for the same spoken language\. We observe a similar trend for X→\\toEn translation on test data that reverses during En→\\toX translation\.

We conjecture that Gemini’s pre\-training exposure to Manipuri in Bengali script is likely far greater than Manipuri in Meitei Mayek \(U\+ABC0–U\+ABFF\), which appears much more scarcely in online text, but that cannot account for the inconsistencies we observe\. In any case, we caution that choice of orthography, where applicable, impacts MT performance using our RAG method\.

#### Data Cleaning\.

In addition to the original grid search described in[Section5](https://arxiv.org/html/2608.13722#S5), we perform the same procedure after the submission deadline using slightly different data and ranking\. We filter the WMT26 training data again, this time removing near duplicate candidate exemplars, and rank on the ChrF\+\+ metric rather than ChrF\.

The results appear in[Table5](https://arxiv.org/html/2608.13722#S7.T5), and comparison with[Table3](https://arxiv.org/html/2608.13722#S6.T3)and[Table4](https://arxiv.org/html/2608.13722#S6.T4)reveals that optimal ther,dr,dconfiguration changes for 18 of 22 tasks\. In particular, the optimalrrfalls to zero in seven out of eleven X→\\toEn translation tasks\. Conversely, nine out of eleven X→\\toEn translation tasks now showd=49d=49\. Six have \(r=0,d=49r=0,d=49\)\. Nine out of eleven En→\\toX tasks still showr=80r=80\.

These results for newly filtered data, although they still show performance for a small pilot ofn=50n=50samples, demonstrate a clear asymmetry with respect to translation direction that our original grid search does not show\. An optimumr=80r=80for most languages in En→\\toX performance suggests that RAG does improve performance while generating a low\-resource language, as in these tasks and our original AmericasNLP shared task submission\. Twor=40r=40configurations appear in Karbi and Nagamese translation, and these feature the smallest corpora, possibly indicating that retrieval strength is limited during extreme resource scarcity\. Meanwhile,r=0r=0in X→\\toEn tasks indicates that retrieval does not help when generating a high\-resource language the LLM already knows well and when there are no samples identical or almost identical to the actual query to draw on\.

#### Source\-Side Retrieval Index\.

For X→\\toEn, building the BM25 index on the Indic language’s side \(rather than English in all settings\) changes the optimal configuration for most languages\. This is an essential improvement to our AmericasNLP 2026 submission[8](https://arxiv.org/html/2608.13722#bib.bib2)\. Under the previous, source\-invariant English BM25 index, the optimalr,dr,dconfiguration hadr=0r=0for As→\\toEn and Kha→\\toEn translation\. The implication is that no retrieval outperforms any retrieval based on the fixed English index for these languages\. When the BM25 index is corrected, the optimalrrincreases to what we present in[Table3](https://arxiv.org/html/2608.13722#S6.T3)and[Table4](https://arxiv.org/html/2608.13722#S6.T4), indicating that retrieval does finally improve translation\.

## 8Discussion

For nearly every language and direction, we observe a some decay in performance from development to test set data\. We expect this decay given our previous results in the AmericasNLP shared task[8](https://arxiv.org/html/2608.13722#bib.bib2)\. The greatest dev\-test differences appear for the Bodo and Kokborok and for En→\\toBodo translation and Trp→\\toEn translation in particular\. We observe a difference of nearly 30 and 40 points, respectively, as we shift from the development to the test data\. Still, Bodo and Kokborok are outliers as compared to the other languages, and there are many cases where test performance is actually greater than dev performance according to the ChrF\+\+ metric\. We build[Figure2](https://arxiv.org/html/2608.13722#S7.F2)in Matplotlib[14](https://arxiv.org/html/2608.13722#bib.bib39)to demonstrate the performance gaps and exceptions to predicted performance decline for both En→\\toX \(right\) and X→\\toEn \(left\) translation tasks\.

For languages where the gap is thinner or flipped, this may a signal that the test dataset for the current shared task is much closer to the same domain as the development set than the datasets in AmericasNLP 2026’s shared task, but a close examination of the test data or thorough documentation of it is necessary to confirm this hypothesis and it would not explain sizable performance gains as we move from development to test data\.

Category 1 languages typically have many times more retrieval pairs than Category 2 languages, Bodo not withstanding\. Least\-squares linear regression analysis with Scipy[44](https://arxiv.org/html/2608.13722#bib.bib40), including Wald t\-tests, reveals no significant \(p<0\.05p<0\.05\) correlation betweenPairsand ChrF\+\+ or BLEU performance, although we do uncover a weak but significant positive correlation \(r=0\.650r=0\.650,p<0\.05p<0\.05,m\>0m\>0\) between augmentation ratio and dev En→\\toX performance\. These findings may be linked to our previous observation that total retrieval bank volume is less important than domain alignment\.

The effect ofrrandddvalues in this shared task is well explored through our grid search\. For Category 2 languages, we observe that the gain fromd=0d=0tod=10d=10exceeds the gain fromr=0r=0tor=80r=80, holding the other hyperparameter constant\. Linear regression analysis also discovers a significant positive correlation between devddvalues and En→\\toX ChrF\+\+ \(r=0\.625r=0\.625,p<0\.05p<0\.05,m\>0m\>0\) but none betweenrrand En→\\toX preformance\. Curated development examplars therefore appear to be a stronger signal than BM25\-retrieved training pairs when the corpus is small\.r=80r=80performs best for ten of the eleven languages in En→\\toX translation\. In particular, Bodo performance peaks at\(r=80,d=49\)\(r=80,d=49\)with a score of 76\.37 ChrF and 72\.14 ChrF\+\+\. We investigate the search grids further and find that the performance gains asrrincreases, holdingddconstant, are somewhat flat for languages with less than 10,000 total exemplars, including Karbi \(899\) and Tagin \(4,421\)\. Karbi is the one language withr<80r<80in our contrastive En→\\toX submissions\. We conclude larger retrieval corpora better reward higher retrieval countsrr\.

## 9Conclusion

We describe the gators submission to WMT26’s Indic Shared Task: A training\-free, retrieval\-augmented system that uses BM25 and Gemini 2\.5 Flash to translate between English and eleven Northeastern Indian languages\. The approach applies transfer learning to our AmericasNLP 2026 system[8](https://arxiv.org/html/2608.13722#bib.bib2)but removes vision to isolate retrieval\-augmented MT and builds retrieval banks with different parallel text\.

Along with other ablations, we complete a grid search across all 22 language–direction pairs, and the submission covers every pair in both directions\. We show the importance of domain matching to the synthetic data augmentation we espouse in the previous shared task[8](https://arxiv.org/html/2608.13722#bib.bib2)and fix a critical bug in our original BM\-25 index construction\.

While the previous shared task submission shows that retrieval\-augmented MT can translate VLM\-generated captions into low\-resource languages, we now isolate the capabilities of retrieval\-augmented MT and reproduce our earlier success\. Our submissions consistently outperform the competition by small yet measurable margins\. In 19 translation tasks, our contrastive submission achieves the best BLEU score for nine and the best ChrF\+\+ score for eleven[28](https://arxiv.org/html/2608.13722#bib.bib8)\.

## Limitations

The system depends entirely on the Gemini 2\.5 Flash API\. During the pilot grid search, burst\-level rate\-limit 429 errors under exponential backoff consumed entire SLURM job windows, requiring careful scheduling and a 2\-key API pool to complete all 22 directions’ grid cells in time for submission\. A small number of test sentences \(4 of 21,599; 0\.018%\) produced empty output due to Gemini’s safety filters blocking specific political content\.

Retrieval bank quality varies widely\. Category 2 languages have limited parallel data, and the available external corpora are often out\-of\-domain\. For instance, our only external corpus for Nagamese consists of Bible translations\. As an ablation in[Section7](https://arxiv.org/html/2608.13722#S7)shows, adding more data does not help when BM25 will not surface it\.

The development split, consisting of the last 500 or fewer pairs of training data and with 50 sentences evaluated per grid cell, shares its domain with the training corpus; so dev metrics overstate performance\. We see a similar effect in AmericasNLP[8](https://arxiv.org/html/2608.13722#bib.bib2)\. In particular, Guaraní dev scores roughly halved during evaluation with the official test set\. This time,[Table3](https://arxiv.org/html/2608.13722#S6.T3)shows an inconsistent drop off for all languages except Bodo, but a performance drop usually exists\.

## Acknowledgments

We thank Arnold and Lisa Goldberg for their financial support and UF Research Computing for HiPerGator access\. We also acknowledge the Gatorade, whose support provided the compute resources necessary for this research\.

## References

- Agrawalet al\.\(2023\)S\. Agrawal, C\. Zhou, M\. Lewis, L\. Zettlemoyer, and M\. GhazvininejadIn\-context examples selection for machine translation\.InFindings of the Association for Computational Linguistics: ACL 2023,Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p3.1)\.
- Brownet al\.\(2020\)T\. Brownet al\.Language models are few\-shot learners\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 1877–1901\.Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p3.1)\.
- Buiet al\.\(2026\)M\. D\. Bui, D\. Guzmán, A\. Ebrahimi, F\. Morales,et al\.Findings of the AmericasNLP 2026 shared task on cultural image captioning for Indigenous languages\.InProceedings of the Sixth Workshop on NLP for Indigenous Languages of the Americas \(AmericasNLP\),San Diego, California\.Cited by:[§1](https://arxiv.org/html/2608.13722#S1.p2.1),[Abstract](https://arxiv.org/html/2608.13722#abstract1.1)\.
- Caswellet al\.\(2025\)I\. Caswell, E\. Nielsen, J\. Luo, C\. Cherry, G\. Kovacs, H\. Shemtov, P\. Talukdar, D\. Tewari, B\. M\. Diane, K\. M\. Doumbouya, D\. Diane, and S\. F\. CisséSMOL: professionally translated parallel data for 115 under\-represented languages\.External Links:2502\.12301,[Link](https://arxiv.org/abs/2502.12301)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px4.p1.1)\.
- Comaniciet al\.\(2025\)G\. Comaniciet al\.Gemini 2\.5: pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities\.arXiv preprint arXiv:2507\.06261\.Cited by:[§1](https://arxiv.org/html/2608.13722#S1.p2.1),[§3\.4](https://arxiv.org/html/2608.13722#S3.SS4.p1.1),[Abstract](https://arxiv.org/html/2608.13722#abstract1.1)\.
- Costa\-Jussàet al\.\(2022\)M\. R\. Costa\-Jussà, J\. Cross, O\. Çelebi,et al\.No language left behind: scaling human\-centered machine translation\.arXiv preprint arXiv:2207\.04672\.Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p2.1),[§7\.1](https://arxiv.org/html/2608.13722#S7.SS1.p2.1)\.
- Dhawanet al\.\(2026a\)A\. Dhawan, C\. Driggers\-Ellis, C\. Grant, and D\. Z\. WangImproving indigenous language machine translation with synthetic data and language\-specific preprocessing\.InProceedings of the Ninth Workshop on Technologies for Machine Translation of Low Resource Languages \(LoResMT 2026\),Rabat, Morocco,pp\. 119–126\.External Links:[Link](https://aclanthology.org/2026.loresmt-1.10/),[Document](https://dx.doi.org/10.18653/v1/2026.loresmt-1.10)Cited by:[§1](https://arxiv.org/html/2608.13722#S1.p2.1),[§7\.1](https://arxiv.org/html/2608.13722#S7.SS1.p1.1)\.
- Dhawanet al\.\(2026b\)A\. Dhawan, C\. Driggers\-Ellis, D\. Kasinets, C\. Grant, and D\. Z\. WangRetrieval\-augmented long\-context translation for cultural image captioning: Gators submission for AmericasNLP 2026 shared task\.InProceedings of the Sixth Workshop on NLP for Indigenous Languages of the Americas \(AmericasNLP\),San Diego, California\.Cited by:[§1](https://arxiv.org/html/2608.13722#S1.p2.1),[§3\.3](https://arxiv.org/html/2608.13722#S3.SS3.p2.1),[§3\.5](https://arxiv.org/html/2608.13722#S3.SS5.p1.1),[§7\.1](https://arxiv.org/html/2608.13722#S7.SS1.p1.1),[§7\.1](https://arxiv.org/html/2608.13722#S7.SS1.p2.1),[§7\.2](https://arxiv.org/html/2608.13722#S7.SS2.SSS0.Px3.p1.1),[§8](https://arxiv.org/html/2608.13722#S8.p1.1),[§9](https://arxiv.org/html/2608.13722#S9.p1.1),[§9](https://arxiv.org/html/2608.13722#S9.p2.1),[Limitations](https://arxiv.org/html/2608.13722#Sx1.p3.1),[§9\.2](https://arxiv.org/html/2608.13722#Sx3.SS2.p1.1),[Abstract](https://arxiv.org/html/2608.13722#abstract1.1)\.
- Driggers\-Elliset al\.\(2025\)C\. Driggers\-Ellis, D\. Brinkley, R\. Chen, A\. Dhawan, D\. Z\. Wang, and C\. GrantMultiScript30k: leveraging multilingual embeddings to extend cross script parallel data\.External Links:2512\.11074,[Link](https://arxiv.org/abs/2512.11074)Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p3.1),[§7\.1](https://arxiv.org/html/2608.13722#S7.SS1.p2.1)\.
- Dugi \(2024\)T\. DugiGinLish v0\.1\.0: first\-ever tagin and english parallel corpus\.Note:NIT Arunachal PradeshDataset, version 0\.1\.0External Links:[Link](https://github.com/repleeka/ginlish)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px4.p1.1)\.
- Elliottet al\.\(2016\)D\. Elliott, S\. Frank, K\. Sima’an, and L\. SpeciaMulti30K: multilingual english\-german image descriptions\.InProceedings of the 5th Workshop on Vision and Language,pp\. 70–74\.External Links:[Document](https://dx.doi.org/10.18653/v1/W16-3210),[Link](http://www.aclweb.org/anthology/W16-3210)Cited by:[§7\.1](https://arxiv.org/html/2608.13722#S7.SS1.p2.1)\.
- Fenget al\.\(2022\)F\. Feng, Y\. Yang, D\. Cer, N\. Arivazhagan, and W\. WangLanguage\-agnostic BERT sentence embedding\.InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px4.p2.1)\.
- Galaet al\.\(2023\)J\. Gala P\. A\. Chitaleet al\.IndicTrans2: towards high\-quality and accessible machine translation models for all 22 scheduled Indian languages\.arXiv preprint arXiv:2305\.16307\.Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p2.1)\.
- Hunter \(2007\)J\. D\. HunterMatplotlib: a 2d graphics environment\.Computing in Science & Engineering9\(3\),pp\. 90–95\.External Links:[Document](https://dx.doi.org/10.1109/MCSE.2007.55)Cited by:[§8](https://arxiv.org/html/2608.13722#S8.p1.1)\.
- Joneset al\.\(2023\)A\. Jones, I\. Caswell, I\. Saxena, and O\. FiratBilex rx: lexical data augmentation for massively multilingual machine translation\.External Links:2303\.15265,[Link](https://arxiv.org/abs/2303.15265)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px4.p1.1)\.
- Joshiet al\.\(2025\)N\. Joshi, P\. Gogoi, A\. Mirza, A\. Jansari, A\. Yadavalli, A\. Pandey, A\. Shukla, D\. Sudharsan, K\. Bali, and V\. SeshadriELR\-1000: a community\-generated dataset for endangered indic indigenous languages\.InProceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Asia\-Pacific Chapter of the Association for Computational Linguistics \(IJCNLP\-AACL 2025\),Mumbai, India\.Note:Preprint: arXiv:2512\.01077External Links:[Link](https://arxiv.org/abs/2512.01077)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px1.p1.1)\.
- Kakumet al\.\(2023\)N\. Kakum, S\. R\. Laskar, K\. Sambyo, and P\. PakrayNeural machine translation for limited resources English–Nyishi pair\.Sādhanā48\(4\)\.Cited by:[§1](https://arxiv.org/html/2608.13722#S1.p1.1)\.
- Karya \(2025\)KaryaELR\-1000: a community\-generated dataset for endangered indic indigenous languages\.Note:Hugging FaceExternal Links:[Link](https://huggingface.co/datasets/karya/ELR-1000)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px1.p1.1)\.
- Karya \(2026\)KaryaEndangered recipes translated 500\.Note:Hugging FaceExternal Links:[Link](https://huggingface.co/datasets/karya/endangered-recipes-translated-500)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px1.p1.1)\.
- Konnerth and Tisso \(2018\)L\. Konnerth and S\. TissoKarbi texts: a fully glossed corpus of different genres\.Himalayan Linguistics17\(2\),pp\. 117–472\.External Links:[Document](https://dx.doi.org/10.5070/H917239461),[Link](https://www.academia.edu/43433609/Karbi_texts_A_fully_glossed_corpus_of_different_genres)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px3.p1.1)\.
- Konnerth \(2014\)L\. KonnerthA grammar of karbi\.Ph\.D\. Thesis,University of Oregon\.External Links:[Link](https://scholarsbank.uoregon.edu/items/5d23a8fb-7937-47a7-8765-7354ce7f542b)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px3.p1.1)\.
- Liuet al\.\(2020\)Y\. Liu, J\. Gu, N\. Goyal, X\. Li, S\. Edunov, M\. Ott, L\. Zettlemoyer, and V\. StoyanovMultilingual denoising pre\-training for neural machine translation\.Transactions of the Association for Computational Linguistics8,pp\. 726–742\.Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p2.1)\.
- Maheshwari and Keshri \(2025\)A\. Maheshwari and P\. KeshriFNPC\-25: first nagamese–english parallel corpus\.Note:KaggleABV\-IIITM Gwalior; supervised by Anjali; 8,771 English–Nagamese parallel sentence pairs; MIT License; accessed July 17, 2026External Links:[Link](https://www.kaggle.com/datasets/belovedorange/first-nagamese-english-parallel-corpus-fnpc-25)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px4.p1.1)\.
- Maitiet al\.\(2025\)A\. Maiti, M\. Pandey, and M\. MandalNagaNLP: bootstrapping NLP for low\-resource Nagamese creole with human\-in\-the\-loop synthetic data\.External Links:2512\.12537,[Link](https://arxiv.org/abs/2512.12537)Cited by:[§4](https://arxiv.org/html/2608.13722#S4.p1.1)\.
- Narzaryet al\.\(2022\)S\. Narzary, M\. Brahma, M\. Narzary, G\. Muchahary, P\. K\. Singh, A\. Senapati, S\. Nandi, and B\. SomGenerating monolingual dataset for low resource language Bodo from old books using Google Keep\.InProceedings of the Thirteenth Language Resources and Evaluation Conference,Marseille, France,pp\. 6563–6570\.External Links:[Link](https://aclanthology.org/2022.lrec-1.705/)Cited by:[§4](https://arxiv.org/html/2608.13722#S4.p1.1)\.
- Pakrayet al\.\(2025\)P\. Pakray, R\. M\. Krishna, S\. Pal, A\. Vetagiri, S\. K\. Dash, A\. K\. Maji, S\. A\. Lyngdoh, L\. Laitonjam, A\. Jamatia, K\. Sambyo, A\. Das, and R\. MannaFindings of WMT 2025 shared task on low\-resource Indic languages translation\.InProceedings of the Tenth Conference on Machine Translation \(WMT\),Suzhou, China,pp\. 532–553\.Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p2.1)\.
- Pakrayet al\.\(2024\)P\. Pakray, S\. Pal, A\. Vetagiri, R\. Krishna, A\. K\. Maji, S\. Dash, L\. Laitonjam, S\. Lyngdoh, and R\. MannaFindings of WMT 2024 shared task on low\-resource Indic languages translation\.InProceedings of the Ninth Conference on Machine Translation \(WMT\),pp\. 654–668\.Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p2.1)\.
- Pakrayet al\.\(28–29\)P\. Pakray, S\. Pal, A\. Vetagiri, K\. B\. Singh, S\. K\. Dash, A\. K\. Maji, S\. A\. Lyngdoh, L\. Laitonjam, A\. Jamatia, A\. Das, S\. Warjri, U\. Sharma, K\. Sambyo, and R\. MannaFindings of WMT shared task on low\-resource Indic languages translation\.InIn Proceedings of the Eleventh Workshop on Machine Translation \(WMT\) under EMNLP 2026,Budapest, Hungary\.Cited by:[§1](https://arxiv.org/html/2608.13722#S1.p1.1),[§2](https://arxiv.org/html/2608.13722#S2.p1.1),[Table 4](https://arxiv.org/html/2608.13722#S6.T4),[§6](https://arxiv.org/html/2608.13722#S6.p2.1),[§9](https://arxiv.org/html/2608.13722#S9.p3.1)\.
- Palet al\.\(2023\)S\. Pal, P\. Pakray, S\. R\. Laskar, L\. Laitonjam, V\. Khenglawt, S\. Warjri, P\. K\. Dadure, and S\. K\. DashFindings of the WMT 2023 shared task on low\-resource Indic language translation\.InProceedings of the Eighth Conference on Machine Translation \(WMT\),pp\. 682–694\.External Links:[Link](https://aclanthology.org/2023.wmt-1.56/)Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p1.1),[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.p1.1)\.
- \[30\]J\. PanjiyarEnglish to bodo parallel dataset\.Note:KaggleLicensed under the GNU Free Documentation License 1\.3External Links:[Link](https://www.kaggle.com/datasets/jigarpanjiyar/english-to-bodo-dataset)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px2.p1.1)\.
- Papineniet al\.\(2002\)K\. Papineni, S\. Roukos, T\. Ward, and W\. ZhuBLEU: a method for automatic evaluation of machine translation\.InProceedings of the 40th Annual Meeting of the Association for Computational Linguistics,pp\. 311–318\.Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p1.1)\.
- Popović \(2015\)M\. PopovićChrF: character n\-gram F\-score for automatic MT evaluation\.InProceedings of the Tenth Workshop on Statistical Machine Translation,Cited by:[§5](https://arxiv.org/html/2608.13722#S5.p1.1)\.
- Popović \(2017\)M\. PopovićChrF\+\+: words helping character n\-grams\.InProceedings of the Second Conference on Machine Translation,Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p1.1),[§5](https://arxiv.org/html/2608.13722#S5.p1.1)\.
- Post and Burling \(2017\)M\. W\. Post and R\. BurlingThe tibeto\-burman languages of northeast india\.Note:InThe Sino\-Tibetan Languages, 2nd ed\., edited by Graham Thurgood and Randy J\. LaPolla, Routledge, pp\. 213–242External Links:[Document](https://dx.doi.org/10.4324/9781315399508),[Link](https://doi.org/10.4324/9781315399508)Cited by:[§4](https://arxiv.org/html/2608.13722#S4.p1.1)\.
- Rameshet al\.\(2022\)G\. Ramesh, S\. Doddapaneni, A\. Bheemaraj, M\. Jobanputra, R\. AK, A\. Sharma, S\. Sahoo, H\. Diddee, J\. Mahalakshmi, D\. Kakwani, N\. Kumar, A\. Pradeep, S\. Nagaraj, K\. Deepak, V\. Raghavan, A\. Kunchukuttan, P\. Kumar, and M\. S\. KhapraSamanantar: the largest publicly available parallel corpora collection for 11 Indic languages\.Transactions of the Association for Computational Linguistics10,pp\. 145–162\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00452)Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px1.p1.1),[Abstract](https://arxiv.org/html/2608.13722#abstract1.1)\.
- Ranathungaet al\.\(2023\)S\. Ranathunga, E\. A\. Lee, M\. Prifti Skenduli, R\. Shekhar, M\. Alam, and R\. KaurNeural machine translation for low\-resource languages: a survey\.ACM Computing Surveys55\(11\),pp\. 1–37\.Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p2.1)\.
- Reiet al\.\(2020\)R\. Rei, C\. Stewart, A\. C\. Farinha, and A\. LavieCOMET: a neural framework for MT evaluation\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p1.1)\.
- Robertson and Zaragoza \(2009\)S\. Robertson and H\. ZaragozaThe probabilistic relevance framework: BM25 and beyond\.Foundations and Trends in Information Retrieval3\(4\),pp\. 333–389\.Cited by:[§3\.3](https://arxiv.org/html/2608.13722#S3.SS3.p1.1),[Abstract](https://arxiv.org/html/2608.13722#abstract1.1)\.
- Sennrichet al\.\(2016\)R\. Sennrich, B\. Haddow, and A\. BirchImproving neural machine translation models with monolingual data\.InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p2.1)\.
- Sharma \(1999\)H\. S\. SharmaA comparison between Khasi and Manipuri word order\.Linguistics of the Tibeto\-Burman Area22\(1\),pp\. 139–148\.External Links:[Document](https://dx.doi.org/10.32655/LTBA.22.1.07),[Link](https://hdl.handle.net/10356/178127)Cited by:[§4](https://arxiv.org/html/2608.13722#S4.p1.1)\.
- Singhet al\.\(2024\)H\. Singh, N\. Gupta, S\. Bharadwaj, D\. Tewari, and P\. TalukdarIndicGenBench: a multilingual benchmark to evaluate generation capabilities of llms on indic languages\.External Links:2404\.16816Cited by:[§3\.1](https://arxiv.org/html/2608.13722#S3.SS1.SSS0.Px1.p1.1)\.
- Singh \(2012\)T\. D\. SinghBidirectional Bengali script and Meetei Mayek transliteration of web\-based Manipuri news corpus\.InProceedings of the 3rd Workshop on South and Southeast Asian Natural Language Processing,Mumbai, India,pp\. 118–123\.External Links:[Link](https://aclanthology.org/W12-5016/)Cited by:[§4](https://arxiv.org/html/2608.13722#S4.p1.1)\.
- Vilaret al\.\(2023\)D\. Vilar, M\. Freitag, C\. Cherry, J\. Luo, V\. Ratnakar, and G\. FosterPrompting PaLM for translation: assessing strategies and performance\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p3.1)\.
- Virtanenet al\.\(2020\)P\. Virtanen, R\. Gommers, T\. E\. Oliphant, M\. Haberland, T\. Reddy, D\. Cournapeau, E\. Burovski, P\. Peterson, W\. Weckesser, J\. Bright, S\. J\. van der Walt, M\. Brett, J\. Wilson, K\. J\. Millman, N\. Mayorov, A\. R\. J\. Nelson, E\. Jones, R\. Kern, E\. Larson, C\. J\. Carey, İ\. Polat, Y\. Feng, E\. W\. Moore, J\. VanderPlas, D\. Laxalde, J\. Perktold, R\. Cimrman, I\. Henriksen, E\. A\. Quintero, C\. R\. Harris, A\. M\. Archibald, A\. H\. Ribeiro, F\. Pedregosa, P\. van Mulbregt, and SciPy 1\.0 ContributorsSciPy 1\.0: Fundamental Algorithms for Scientific Computing in Python\.Nature Methods17,pp\. 261–272\.External Links:[Document](https://dx.doi.org/10.1038/s41592-019-0686-2)Cited by:[§8](https://arxiv.org/html/2608.13722#S8.p3.1)\.
- Xiaoet al\.\(2025\)B\. Xiao, Q\. Shen, and D\. Z\. WangFrom text to multi\-modal: advancing low\-resource\-language translation through synthetic data generation and cross\-modal alignments\.InProceedings of the Eighth Workshop on Technologies for Machine Translation of Low\-Resource Languages \(LoResMT 2025\),Albuquerque, New Mexico, U\.S\.A\.,pp\. 24–35\.External Links:[Link](https://aclanthology.org/2025.loresmt-1.4/),[Document](https://dx.doi.org/10.18653/v1/2025.loresmt-1.4)Cited by:[§2](https://arxiv.org/html/2608.13722#S2.p3.1)\.

## Appendix

This final supplementary section is dedicated to communicating fine\-grained details of our submissions that are not in focus in the main body of our paper\. We summarize in[Section9\.1](https://arxiv.org/html/2608.13722#Sx3.SS1)the prompt engineering that goes into calling Gemini 2\.5 Flash during each experiment described and during testing for each language and translation direction\.[Section9\.2](https://arxiv.org/html/2608.13722#Sx3.SS2)discusses ther,dr,dgrid search we conduct over a pilot sample of the available WMT26 data to select retrieval hyperparameter configurations for our final submission\.

### 9\.1Prompts

In[Table6](https://arxiv.org/html/2608.13722#Sx3.T6)and[Table7](https://arxiv.org/html/2608.13722#Sx3.T7), we describe the considerations taken at the prompting stage for each language and translation direction\. In general, we conditioned Gemini 2\.5 Flash to generate the correct script and encoding during En→\\toX translation through our prompts\. We also inform Gemini about morphological considerations including word order, tonality, and type of agglutination in the Indic target language where applicable\. Because Gemini is asked to generate En during X→\\toEn translation, we find it much less necessary to precondition generation in the high\-resource language through extensive prompting\.

### 9\.2r,dr,dSearch Grid

Similar to our AmericasNLP 2026 shared task submission[8](https://arxiv.org/html/2608.13722#bib.bib2), we define a search grid consisting of all\(r,d\)∈\{0,20,40,80\}×\{0,10,20,49\}\(r,d\)\\in\\\{0,20,40,80\\\}\\times\\\{0,10,20,49\\\}and record ChrF and ChrF\+\+ metrics for each tuple\(r,d\)\(r,d\)during an evaluation with a pilot set ofn=50n=50WMT26 samples\. We designate ChrF as the primary metric for the primary grid search, and our submission for each language and translation direction has the optimalr,dr,dhyperparameter configuration we determine through this grid search\. After the submission deadline, we repeat the procedure with deduplicated samples and rank according to the ChrF\+\+ metric\.

Inclusion in this appendix of a full search over the eleven shared task languages in both translation directions, a total of 22 translation tasks, would imply tabulating 22×\\times16 or 352 grid search cells in the following pages\. This is doubled to 704 cells if we include the second search’s results\. Instead, we refer the reader to our Github page, where we host a[bash script](https://github.com/dhawan98/Gators_wmt26/blob/main/slurm_pilot_grid.sh)that can recreate our grid search experiments using WMT26 data\.

Table 6:We describe the prompting strategies we utilize for each Category 1 language in this table\. Links in this table direct the reader to each prompt where it lives in our public Github repository\.Table 7:We describe the prompting strategies we utilize for each Category 2 language in this table\. Links in this table direct the reader to each prompt where it lives in our public Github repository\.