SAGE: Sustainable Agent-Guided Expert-tuning for Culturally Attuned Translation in Low-Resource Southeast Asia
Authors: Zhixiang Lu, Chong Zhang, Yulong Li, Angelos Stefanidis, Anh Nguyen, Imran Razzak, Jionglong Su, Zhengyong Jiang
Organizations: Xi’an Jiaotong-Liverpool University Suzhou, China · University of Liverpool Liverpool, United Kingdom · Mohamed bin Zayed University of Artificial Intelligence Abu Dhabi, United Arab Emirates
The vision of an inclusive World Wide Web is impeded by a severe linguistic divide, particularly for communities in low-resource regions of Southeast Asia. While large language models (LLMs) offer a potential solution for translation, their deployment in data-poor contexts faces a dual challenge: the scarcity of high-quality, culturally relevant data and the prohibitive energy costs of training on massive, noisy web corpora. To resolve the tension between digital inclusion and environmental sustainability, we introduce Sustainable Agent-Guided Expert-tuning (SAGE). This framework pioneers an energy-aware paradigm that prioritizes the "right data" over "big data". Instead of carbon-intensive training on unfiltered datasets, SAGE employs a reinforcement learning (RL) agent, optimized via Group Relative Policy Optimization (GRPO), to autonomously curate a compact training set. The agent utilizes a semantic reward signal derived from a small, expert-constructed set of community dialogues to filter out noise and cultural misalignment. We then efficiently fine-tune open-source LLMs on this curated data using Low-Rank Adaptation (LoRA). We applied SAGE to translation tasks between English and seven low-resource languages (LRLs) in Southeast Asia. Our approach establishes new state-of-the-art performance on BLEU-4 and COMET-22 metrics, effectively capturing local linguistic nuances. Crucially, SAGE surpasses baselines trained on full datasets while reducing data usage by 97.1% and training energy consumption by 95.2%. By delivering high-performance models with a minimal environmental footprint, SAGE offers a scalable and responsible pathway to bridge the digital divide in the Global South.
Figures & tables
Figure 1. Architectural overview of the SAGE framework.
Figure 2. The SAGE training and alignment pipeline. Stage 1 employs a GRPO-optimized RL agent to curate a subset Dcur from a noisy pool Dnoisy , guided by semantic proximity to a small expert reference Dexp . Stage 2 utilizes LoRA to efficiently fine-tune the LLM on Dcur , minimizing computational overhead while maximizing cultural alignment.
Model
BLEU-4 ↑
COMET-22 ↑
Avg. Tok. ↓
bn
fil
hi
km
lo
my
vi
bn
fil
hi
km
lo
my
vi
Closed-Source Models
GPT-4o
40.15
45.88
42.50
35.10
28.33
31.05
45.13
83.50
84.10
84.05
81.90
79.25
80.11
85.55
92.50
Claude-3.5 Sonnet
41.25
45.15
42.18
34.95
32.12
33.85
45.24
84.10
83.95
83.99
81.75
80.14
82.45
85.79
95.10
Grok-3
41.50
46.10
43.85
36.20
33.55
32.90
46.15
84.10
84.90
84.95
82.88
81.50
81.75
86.10
93.40
Gemini-2.5 pro
38.20
42.55
40.13
32.80
26.57
29.88
42.01
82.15
82.50
82.88
80.50
78.89
79.50
84.60
90.20
Table 1. Comparison for English-to-Southeast-Asian translation in LMICs. The best result is in bold , the second best is underlined . Avg. Tok. denotes estimated average inference token consumption per sample.
Configuration
Data Used
BLEU-4 ↑
Emission
bn
hi
vi
fil
my
km
lo
Avg.
CO2 (kg) ↓
Full SAGE Framework
3% (Curated)
47.15
48.80
48.95
48.55
33.15
41.50
37.10
43.60
4.2 (-95.1%)
- w/o Expert Reward
3% (Heuristic)
41.50
44.15
44.15
45.20
28.50
36.88
31.50
38.84
4.1 (-95.2%)
- w/o RL Curation
3% (Random)
35.10
39.50
39.50
40.20
23.15
33.50
25.13
33.73
3.8 (-95.6%)
Baseline (Full Dataset)
100% (Noisy)
33.05
38.15
38.05
39.10
22.50
32.10
22.05
32.14
85.6 (Ref.) †
Table 2. Ablation study of the SAGE framework on the Qwen-3-8B model. We analyze the impact of removing key components (w/o) on BLEU-4 scores across 7 languages. The rightmost column demonstrates the environmental efficiency of SAGE compared to the baseline.
Training Scheme
Performance Metric ↑
Sig.
BLEU-4
COMET-22
No-Filter (Baseline)
21.5
68.2
–
BLEU-Reward RL ( Sommerschield et al., 2023 )
32.8 (+52.6%)
74.5 (+9.2%)
∗
QE-based Filtering ( Zhang et al., 2020 )
33.9 (+57.7%)
76.1 (+11.6%)
‡
SAGE Framework (ours)
38.5 (+79.1%)
81.4 (+19.4%)
‡
Table 3. Comparison of data curation schemes across five LRLs. The “Sig.” column denotes statistical significance compared to the No-Filter baseline ( ‡ : p<0.01 , ∗ : p<0.05 ).
Language
Mean BLEU-4 Δ
p-value
Bengali (bn)
9.65
0.0021
Filipino (fil)
4.70
0.0183
Hindi (hi)
9.25
0.0028
Khmer (km)
9.35
0.0025
Lao (lo)
7.70
0.0090
Burmese (my)
5.65
0.0112
Table 4. Statistical significance of BLEU-4 improvements from applying our SAGE framework to the Qwen3 8B base model. P-values from a paired t-test confirm that all improvements are statistically significant (p < 0.05).
Figure 3. Sensitivity analysis of performance relative to expert data size ( ∣Dexpert∣ ). The dashed black line represents the average BLEU-4 score across all seven languages.
Dexpert Size
Human Cost
Avg. BLEU-4
Sig.
0 (Baseline)
0.0 hrs
38.84
–
100 Pairs
1.1 hrs
41.22 (+6.1%)
500 Pairs
5.4 hrs
42.89 (+10.4%)
1,000 Pairs
10.8 hrs
43.36 (+11.6%)
2,000 Pairs
21.6 hrs
43.60 (+12.3%)
Table 5. Sensitivity Analysis of Expert Set Size ( Dexpert ). Results are averaged across 7 languages. “Human Cost” denotes the estimated annotation time (approx. 1 min/pair). Improvements are relative to the Baseline. Significance: * ( p<0.05 ).
Source (EN)
"You should take this medicine after meals to avoid stomach pain". (Context: Doctor to elderly patient)
Baseline
"Bạn nên uống thuốc này sau bũ\kern-0.46252pt\raise0.77498pt\hbox{'{}}a ăn..".
(Analysis)
Generic pronoun "Bạn" (friend) is inappropriate/rude for an elder.
SAGE
" Bác nên uống thuốc này sau bũ\kern-0.46252pt\raise0.77498pt\hbox{'{}}a ăn..".
(Analysis)
Honorific "Bác" (Uncle/Elder) correctly reflects social hierarchy.
Reference
"Bác nên dùng thuốc này sau khi ăn..".
Table 6. Case study on cultural subtleties of community dialogues.
Figure 4. Efficiency evaluation of the SAGE framework. (a) Environmental Efficiency: SAGE reduces carbon emissions by over 95% compared to baseline fine-tuning by leveraging high-quality, culturally attuned data subsets. The hatched area represents the carbon savings. (b) Inference Throughput: Comparison of token generation speed.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Language
Annotators
Avg. Speed
Total Cost
(Count)
(Pairs/Hour)
(2k Pairs)
Bengali (bn)
2
100
20.0 hrs
Filipino (fil)
2
110
18.2 hrs
Hindi (hi)
2
95
21.1 hrs
Khmer (km)
2
80
25.0 hrs
Lao (lo)
2
85
23.5 hrs
Appendix
Table 7. Annotation Cost Breakdown per Language. Details the number of expert annotators and the specific time consumed to construct the full Dexpert (2,000 pairs). Speed indicates the average sentence pairs translated per hour.
Language
0 Pairs
100 Pairs
500 Pairs
1,000 Pairs
2,000 Pairs
bn
42.00
44.58
46.38
46.89
47.15
fil
43.25
45.90
47.75
48.28
48.55
hi
43.47
46.14
48.00
48.53
48.80
km
36.97
39.23
40.82
41.27
41.50
lo
33.05
35.07
36.49
36.90
37.10
my
29.53
31.34
32.61
32.97
33.15
Appendix
Table 8. Detailed breakdown of BLEU-4 scores by language and Expert Set Size.
Model
Dtype *
Mem.
Speed
BLEU-4
COMET
(GB)
(Tok/s)
( ↑ )
( ↑ )
Baselines
NLLB-200-3.3B
FP16
7.2
76.5
25.80
76.50
Llama-3.1-8B
FP16
OOM †
–
34.50 ‡
78.20 ‡
Gemma-3-9B
FP16
OOM †
–
35.10 ‡
79.10 ‡
SAGE (Ours)
Appendix
Table 9. Inference efficiency and quality comparison on NVIDIA T4 (16GB). Comparison of resource usage and translation quality.
Figure 5. Environmental Impact: SAGE achieves comparable or superior performance while reducing training data usage by 97% and carbon footprint by over 95% compared to standard fine-tuning.
Figure 6. The landscape of selected Asian languages, plotted by speaker population against the economic level of their primary country. Bubble size is proportional to the speaker population.
Machine translation develops computational methods for converting text, speech, and visual content between languages while preserving meaning and purpose.
Large Language Models (LLMs) offer new potential for translation tasks but often experience performance degradation when handling low-resource languages. To address this limitation, we propose an approach for fine-tuning LLMs on a low-resource language, Kupang Malay. Our approach involves designing a set of instructions by leveraging explicit lexical and semantic features from a bilingual dictionary, and introducing Continual Instruction Tuning (CIT), a training paradigm that enables iterative instruction-based training. Experimental results demonstrate that our model, named Lius, yields notable improvements over standard instruction-tuned models by outperforming 4-6 points, and surpassing both Neural Machine Translation (NMT) and Multilingual LLM models by 10-13 points on several evaluation metrics. These findings highlight the potential of our approach to mitigate the reliance on large-scale parallel data in low-resource language translation.
Neural machine translation (NMT) from Chinese to low-resource Southeast Asian languages remains severely constrained by the extreme scarcity of clean parallel corpora and the pervasive noise in existing mined data. This chronic shortage not only impedes effective model training but also sustains a large performance gap with high-resource directions, leaving millions of speakers of languages such as Lao, Burmese, and Tagalog with persistently low-quality translation systems despite recent advances in large multilingual models. We introduce Multilingual Expert-Reward Informed Tuning (MERIT), a unified translation framework that transforms the traditional English-centric ALT benchmark into a Chinese-centric evaluation suite for five Southeast Asian low-resource languages (LRLs). Our framework combines language-specific token prefixing (LTP) with supervised fine-tuning (SFT) and a novel group relative policy optimization (GRPO) guided by the semantic alignment reward (SAR). These results confirm that, in LRL{\textrightarrow}Chinese translation, targeted data curation and reward-guided optimization dramatically outperform mere model scaling.
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.