Can Linguistically Related Languages Guide LLM Translation in Low-Resource Settings?

Aishwarya Ramasethu; Niyathi Allu; Rohin Garg; Harshwardhan Fartale; Dun Li Chan

cs.CLMar 17, 2026

Can Linguistically Related Languages Guide LLM Translation in Low-Resource Settings?

Authors: Aishwarya Ramasethu, Niyathi Allu, Rohin Garg, Harshwardhan Fartale, Dun Li Chan

Organizations: Prediction Guard · Independent · Scale AI · INTI International College Penang

Abstract

Large Language Models (LLMs) have achieved strong performance across many downstream tasks, yet their effectiveness in extremely low-resource machine translation remains limited. Standard adaptation techniques typically rely on large-scale parallel data or extensive fine-tuning, which are infeasible for the long tail of underrepresented languages. In this work, we investigate a more constrained question: in data-scarce settings, to what extent can linguistically similar pivot languages and few-shot demonstrations provide useful guidance for on-the-fly adaptation in LLMs? We study a data-efficient experimental setup that combines linguistically related pivot languages with few-shot in-context examples, without any parameter updates, and evaluate translation behavior under controlled conditions. Our analysis shows that while pivot-based prompting can yield improvements in certain configurations, particularly in settings where the target language is less well represented in the model's vocabulary, the gains are often modest and sensitive to few shot example construction. For closely related or better represented varieties, we observe diminishing or inconsistent gains. Our findings provide empirical guidance on how and when inference-time prompting and pivot-based examples can be used as a lightweight alternative to fine-tuning in low-resource translation settings.

Figures & tables

Figure 1: The Training Data Structure. The input $x$ contains the system instruction, retrieved semantic context (top-5), and the pivot scaffolding. The target $y$ contains only the model’s generated translation.

Table 2

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

			BLEU		chrF++
Lang.	Mod.	$k$	$\Delta$	$p$	$\Delta$	$p$
Gom	Her	$0$	$+0.12$	$.08$	$+0.22$	$.20$
	Her	$1$	$-0.18$	$.88$	$-0.89$	$1.0$
	Her	$2$	$-0.12$	$.75$	$-0.33$	$.89$
	Tow	$1$	$-0.02$	$.57$	$-0.84$	$1.0$
	Tow	$2$	$+0.07$	$.23$	$-1.11$	$1.0$

Appendix

Table 3: Paired bootstrap significance (pivot

-

direct). No comparison reaches

p<0.05

. Gom: Konkani (

n{=}205

), Aeb: Tunisian Arabic (

n{=}100

). Her: Hermes, Tow: Tower.

Metric	Signature
BLEU	nrefs:1\|case:mixed\|eff:no\|
	tok:13a\|smooth:exp\|version:2.5.1
chrF++	nrefs:1\|case:mixed\|eff:yes\|
	nc:6\|nw:2\|space:no\|version:2.5.1

Appendix

Table 4: SacreBLEU scoring signatures.

Language Pair	BLEU	chrF++
Eng-Gom	7.51	26.82
Eng-Aeb	4.20	10.42

Appendix

Table 5: NLLB-200 distilled reference baseline results for our evaluation datasets.

Dataset	Language	Tower	Hermes
Gom	Eng	1.59	1.34
Gom	Mar	7.73	4.08
Gom	Gom	7.65	4.09
Aeb	Eng	1.27	1.21
Aeb	MSA	4.74	2.12
Aeb	Aeb	4.96	2.16

Appendix

Table 6: Tokens per word across languages for Tower and Hermes models. Lower values indicate more efficient tokenization.

Language	Model	k=1	k=2	k=3	k=4	k=5
Arabic	Hermes	24.09	24.89	25.17	23.79	24.06
Arabic	Tower	13.08	12.70	11.85	12.06	11.68
Konkani	Hermes	28.96	27.45	27.82	25.95	26.80
Konkani	Tower	10.78	8.71	9.69	9.16	8.24

Appendix

Table 7: chrF scores between pivot translations and generated translations for different values of

k

. Lower scores indicate greater divergence from the pivot output, suggesting that the model is not simply reproducing the pivot translations.

Language Pair	Jaccard Similarity
Eng-Mar	0.0002
Eng-Gom	0.0121
Mar-Gom	0.1054

Appendix

Table 8: Word-level Jaccard similarity scores for the Konkani corpus. Marathi and Konkani show substantially higher lexical overlap than English with either language.

Language Pair	Jaccard Similarity
Eng-MSA	0.0010
Eng-Aeb	0.0010
MSA-Aeb	0.1646

Appendix

Table 9: Word-level Jaccard similarity scores for the Arabic corpus. MSA and Tunisian Arabic show moderate lexical overlap, while both exhibit minimal overlap with English.

Model	Source	Target	$k$	BLEU	chrF++
Ablation: Number of In-Context Examples ( $k$ )
Unbabel/TowerInstruct-v0.1
Unbabel/TowerInstruct-v0.1	Eng	Gom	0	1.28	0.69
Unbabel/TowerInstruct-v0.1	Eng	Gom	1	3.67	21.01
Unbabel/TowerInstruct-v0.1	Eng	Gom	2	3.38	21.25
Unbabel/TowerInstruct-v0.1	Eng	Gom	3	3.39	19.30

Appendix

Table 10: Ablation on the number of in-context examples (

k

) for English

\rightarrow

Konkani direct translation.

Model	Source	Target	$k$	BLEU	chrF++
Ablation: Number of In-Context Examples ( $k$ )
Unbabel/TowerInstruct-v0.1
Unbabel/TowerInstruct-v0.1	Eng	Aeb	0	4.19	17.62
Unbabel/TowerInstruct-v0.1	Eng	Aeb	1	4.46	19.49
Unbabel/TowerInstruct-v0.1	Eng	Aeb	2	4.46	16.23
Unbabel/TowerInstruct-v0.1	Eng	Aeb	3	4.07	15.59

Appendix

Table 11: Ablation on the number of in-context examples (

k

) for English

\rightarrow

Tn direct translation.

Model	Source	Pivot	Target	$k$	BLEU	chrF++
Ablation: Number of In-Context Examples ( $k$ ) using Marathi Pivot (No Fine-Tuning)
Unbabel/TowerInstruct-v0.1
Unbabel/TowerInstruct-v0.1	English	Mar	Gom	0	2.07	1.30
Unbabel/TowerInstruct-v0.1	English	Mar	Gom	1	2.58	16.03
Unbabel/TowerInstruct-v0.1	English	Mar	Gom	2	5.68	8.94
Unbabel/TowerInstruct-v0.1	English	Mar	Gom	3	4.11	17.66

Appendix

Table 12: Ablation on the number of in-context examples (

k

) for English

\rightarrow

Marathi

\rightarrow

Konkani translation.

Model	Source	Pivot	Target	$k$	BLEU	chrF++
Ablation: Number of In-Context Examples ( $k$ ) using MSA Pivot (No Fine-Tuning)
Unbabel/TowerInstruct-v0.1
Unbabel/TowerInstruct-v0.1	Eng	Msa	Aeb	0	4.37	16.45
Unbabel/TowerInstruct-v0.1	Eng	Msa	Aeb	1	3.46	18.74
Unbabel/TowerInstruct-v0.1	Eng	Msa	Aeb	2	4.99	16.57
Unbabel/TowerInstruct-v0.1	Eng	Msa	Aeb	3	4.77	17.32

Appendix

Table 13: Ablation on the number of in-context examples (

k

) for English

\rightarrow

Msa

\rightarrow

Aeb translation

Model	Source	Pivot	Target	$k$	BLEU	chrF++	$\Delta$ chrF++
Ablation: Number of In-Context Examples ( $k$ ) using Hindi Pivot (No Fine-Tuning)
NousResearch/Hermes-2-Pro-Llama-3-8B
NousResearch/Hermes-2-Pro-Llama-3-8B	English	Hin	Gom	0	2.86	25.39	+0.49
NousResearch/Hermes-2-Pro-Llama-3-8B	English	Hin	Gom	1	2.47	24.12	-6.22
NousResearch/Hermes-2-Pro-Llama-3-8B	English	Hin	Gom	2	2.47	23.96	-3.63
NousResearch/Hermes-2-Pro-Llama-3-8B	English	Hin	Gom	3	2.41	23.69	-2.20

Appendix

Table 14: Ablation on the number of in-context examples (

k

) for English

\rightarrow

Hindi

\rightarrow

Konkani translation using Nous Hermes.

\Delta

chrF++ is computed relative to the Marathi-pivot setting at the same

k

. Scores computed with SacreBLEU ( Post, 2018 ) ; signatures in Appendix A.2 .

Parameter	Value
batch_size	1
num_train_epochs	1.5
warmup_ratio	0.03
logging_steps	25
learning_rate	2e-4
gradient_checkpointing	True

Appendix

Table 15: Training parameters used in the model training process.

Parameter	Value
r	64
lora_alpha	16
lora_dropout	0.1
bias	none
task_type	CAUSAL_LM
target_modules	[’q’, ’k’, ’v’, ’o’, ’up’, ’down’, ’gate’, ’lm_head’]

Appendix

Table 16: LoRA configuration parameters.

do_sample : True

temperature : 0.1

num_return_sequences : 1

max_new_tokens : 200

return_full_text : False

Appendix

Table 17: Inference parameters used for text generation.

Model	Source	Pivot	Target	BLEU	CHRF++
Few-shot Finetuned
Unbabel/TowerInstruct-v0.1	English	-	Konkani	4.18	31.57
NousResearch/Hermes-2-Pro-Llama-3-8B	English	-	Konkani	3.49	31.49
Unbabel/TowerInstruct-v0.1	English	Marathi	Konkani	7.80	17.60
NousResearch/Hermes-2-Pro-Llama-3-8B	English	Marathi	Konkani	12.14	34.92
Zero-shot Finetuned

Appendix

Table 18: Performance comparison of finetuned models in few-shot and zero-shot settings for Konkani translation.

Model	Source	Pivot	Target	BLEU	CHRF++
Few-shot Finetuned
Unbabel/TowerInstruct-v0.1	English	-	Tn	3.3	21.05
NousResearch/Hermes-2-Pro-Llama-3-8B	English	-	Tn	NA	NA
Unbabel/TowerInstruct-v0.1	English	Msa	Tn	2.82	17.12
NousResearch/Hermes-2-Pro-Llama-3-8B	English	Msa	Tn	8.02	35.99
Zero-shot Finetuned

Appendix

Table 19: Performance comparison of finetuned models in few-shot and zero-shot settings for Tunisian Arabic translation.

Explore similar work

Back-Translation Still Matters: Leveraging Monolingual Data for Low-Resource Machine Translation with LLMs
Oct 13, 2026Divyanshu Bathla, Parag Singla, Jitin SinglaLow-Resource MT Machine Translation
IIT Roorkee · IIT Delhi
Left Behind: Cross-Lingual Transfer as a Bridge for Low-Resource Languages in Large Language Models
Mar 22, 2026Abdul-Salem BeibitkhanLow-Resource Language Processing Multilingual Language Model Evaluation
North American University
Reasoning over Grammar: Can Synthetic Linguistic Reasoning Traces Enhance Low-Resource Machine Translation?
Jun 2, 2026Renhao Pei, Yihong Liu, Sampo Pyysalo +2CoT Reasoning Low-Resource MT
ELLIS Institute Finland · University of Turku · Center for Information and Language Processing, LMU Munich +1

Oct 13, 2026J/K move · Enter open · S save

Back-Translation Still Matters: Leveraging Monolingual Data for Low-Resource Machine Translation with LLMs

Divyanshu Bathla, Parag Singla, Jitin Singla

IIT Roorkee · IIT Delhi

Low-Resource MT Machine Translation LLM Fine-Tuning Self-Training for Language Models

Open paper AskPDF arXiv

Pre-trained Large Language Model (LLMs) are effective on many tasks in zero-shot and few-shot settings. For machine translation of low-resource languages, the obvious options are few-shot prompting and fine-tuning on a small parallel set. We ask: what if moderate amounts of low-resource monolingual data are also available? We argue that, rather than continued pretraining, a more effective use of this data is the pre-LLM era technique of back-translation (BT). The current fine-tuned LLM translates monolingual sentences in each language, and the resulting pairs train it in the reverse direction. Updates on the parallel set are interleaved for stable learning. With 1K parallel pairs on Odia and Swahili, BT beats continued pretraining into the low-resource language by 7.7 to 16.4 XCOMET-XXL points in domain. With beam-decoded intermediate translations, it also beats a model with 10 time more parallel data. In-context examples retrieved from the parallel data further improve BT. We also analyze the source, quality and amount of monolingual text. Across five low-resource languages, with Hindi as a high-resource control, and two 3B backbones, BT improves translation into the low-resource language by 16 to 173% XCOMET-XXL over parallel-only fine-tuning. Blind human evaluation on Odia and Sanskrit agrees with the automatic metrics.

Abstract

Explore similar work

Back-Translation Still Matters: Leveraging Monolingual Data for Low-Resource Machine Translation with LLMs

Left Behind: Cross-Lingual Transfer as a Bridge for Low-Resource Languages in Large Language Models

Reasoning over Grammar: Can Synthetic Linguistic Reasoning Traces Enhance Low-Resource Machine Translation?