As globalization accelerates the demand for seamless cross-border enterprise collaboration, multilingual translation remains bottlenecked by insufficient domain-specific terminology, rigid terminology translation, weak instruction-following capabilities, and high inference latency in long-instruction scenarios. To overcome these barriers, we propose Bridging the Terminology Gap (BTG), a data-centric framework. First, BTG mines cross-lingual term alignments from domain documents and leverages LLMs to synthesize terminology-rich parallel corpora. Second, a hybrid filtering strategy combining heuristics and XCOMET scoring ensures high data fidelity. Third, a two-stage training paradigm employs supervised fine-tuning (SFT) for targeted terminology injection, followed by reinforcement learning (RL) to refine output preferences, effectively balancing domain knowledge integration with model reliability. Extensive offline evaluations (e.g., XCOMET, GEMBA, Acc@T) and online human assessments demonstrate significant improvements in terminology accuracy and instruction-following capabilities, validating the effectiveness of our framework, which is now deployed on our cross-border collaboration platform.
Figures & tables
Figure 1: Illustration of critical challenges in cross-border collaboration translation.
Figure 2: Overall architecture of the BTG framework: data mining and augmentation(left), data filtering and mixing (middle), model training and evaluation(right).
Table 1: Performance of BTG-tuned Qwen3-8B vs. 9 baselines across 10 metrics (GM: GEMBA, CM: COMET, XCM: XCOMET).
Method
Acc@T
Avg. Latency(s)
Chain-of-Dictionary
44.76
4.73
Translate-and-Revise
83.92
6.26
DeepTrans
41.24
3.10
MT-R1-Zero
40.43
3.79
Ours(BTG)
93.97
1.70
Table 2: Term accuracy and latency comparison against SOTA methods.
Figure 3: Ablation study on the effect of data quality.
Figure 4: Impact of data composition:(top) performance across General:Term:Instruction ratios, and (bottom) Acc@T scaling with general data under fixed term.
Figure 5: Expert evaluation of Best Preference and Acceptable Candidates on online documents.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
General Trans. Domain
Terminology Domain
Zh-En-Pt Parallel Chat
1,676,758
Internet & Corp. Term.
229,718
United Nations Docs
1,084,881
Instruction Following
Opus-100
643,610
Orca
994,896
Literature and Film
329,077
Belle
519,255
Religion
7,270
Alpaca
51,760
Corp. Encyclopedia
4,664
Appendix
Table 3: Detailed statistics of training datasets.
Figure 6: Performance trends under LLM-Driven data mixtures.
Table 4: LLM-Driven mode system prompt
Table 5: LLM-Driven mode 1st round
Table 6: LLM-Driven mode 2rd round
Table 7: LLM-Driven mode 3rd round
Table 8: LLM-Driven mode 4th round and final result
Table 10: An example of extracting practical terms, extracting Chinese “步兵” and English "the on foot" from similar sentences
Domain
Mined
Aug.
Train
Acc@T
Pairs
Pairs
Set
Biomedicine
828
8,073
40K
41.84 → 70.41
Agriculture
295
2,841
20K
43.77 → 64.33
Appendix
Table 11: Statistics and performance across specialized domains. The significant gains in Acc@T demonstrate cross-domain generalizability.
Model
CMKiwi
XCM-XL
XCM-ref
Acc@T
English → Portuguese
Qwen3-8B-BTG
84.56
92.82
94.43
95.36
Qwen3-8B
83.27
86.69
86.74
50.63
Qwen2.5-7B
79.72
80.40
80.54
41.65
Qwen3-Max
81.99
83.68
84.69
50.98
Qwen-MT
84.24
88.49
88.99
54.34
Appendix
Table 12: Cross-lingual breakdown analysis of the main experiment. Top: English to Portuguese. Bottom: Portuguese to Chinese. (CM: COMET)
Length
TTFT (s)
Gen Time (s)
Hallu. (%)
(Tokens)
BTG
Base
BTG
Base
BTG
Base
0–1.5k
0.41
11.66
12.26
12.32
3.96
9.63
1.5k–3k
0.50
13.78
21.74
21.22
6.34
10.47
3k–5k
0.63
16.39
30.43
29.24
5.96
12.08
Appendix
Table 13: Inference efficiency comparison across different input lengths (Hallu:Hallucination Rate ).
Figure 7: Impact of training methods on model performance.
Figure 8: Performance of trained vs. base models on 4 tasks: correction, structured output, style transfer, and custom terminology.
Table 14: A task example of user custom term injection
Table 15: An example of the terminology correction task.
Table 16: An example of structured output translation task.
Table 17: An example of sytle transfer translation task.
Machine translation develops computational methods for converting text, speech, and visual content between languages while preserving meaning and purpose.
Multilingual Large Language Models (LLMs) struggle with cross-lingual tasks due to data imbalances between high-resource and low-resource languages, as well as monolingual bias in pre-training. Existing methods, such as bilingual fine-tuning and contrastive alignment, can improve cross-lingual performance, but they often require extensive parallel data or suffer from instability. To address these challenges, we introduce a Cross-Lingual Mapping Task during the pre-training phase, which enhances cross-lingual alignment without compromising monolingual fluency. Our approach bi-directionally maps languages within the LLM embedding space, improving both language generation and comprehension. We further propose a Language Alignment Coefficient to robustly quantify cross-lingual consistency, even in limited-data scenarios. Experimental results on machine translation (MT), cross-lingual natural language understanding (CLNLU), and cross-lingual question answering (CLQA) show that our model achieves gains of up to 11.9 BLEU points in MT, 6.72 points in CLQA BERTScore-Precision, and more than 5% in CLNLU accuracy over strong multilingual baselines. These findings highlight the potential of incorporating cross-lingual objectives into pre-training to improve multilingual LLMs.
Weihua Zheng, Chang Liu, Zhengyuan Liu +5
Singapore University of Technology and Design, Singapore · ByteDance, Singapore · Agency for Science, Technology and Research, Singapore
Modern translation workflows demand more than semantic equivalence. Users routinely require models to preserve JSON or HTML schemas, honor curated glossaries, disambiguate with provided context, and match prescribed registers, often several at once. Conventional metrics such as BLEU and xCOMET capture semantic fidelity but provide little signal on constraint adherence, while general instruction following benchmarks ignore the cross-lingual nature of translation. We introduce \bench, a benchmark for multilingual translation instruction following covering seven languages, with 4,506 single-constraint and 2,838 multi-constraint items spanning six constraint dimensions and five compositional patterns with instructions issued in all seven languages. Constraints are split into a gating subset verified by deterministic checkers and a continuous subset scored by a rubric-based LLM judge, combined under a multiplicative rule that resists reward hacking. Evaluating 15 models reveals systematic gaps that prior protocols miss: Instruction following scales with size more sharply than translation quality, glossary and structured-format constraints dominate the difficulty gradient, and general instruction following rankings correlate only weakly with translation behavior. Our benchmark are available at https://github.com/Tencent-Hunyuan/Hy-MT2/tree/main/IFMTBench.
Terminology-aware translation asks for more than a correct translation: the output must use the exact terms a glossary prescribes. The standard recipe, fine-tuning on glossary-annotated translation pairs, hides an inefficiency: for most examples the glossary prescribes exactly what the model would have produced anyway, so they teach nothing about following a glossary. We therefore keep only the examples where the model's own translation contradicts the glossary. In a controlled study at fixed data volume, this selection alone raises term accuracy from 78.7% to 89.9%. The filtered data, built by a two-way synthetic pipeline on open models, is part of the instruction-tuning mixture of our public release SalamandraTA-7b-instruct v3.0, which, used exactly as released and wrapped in a document-level inference pipeline, forms the BSC submission to the WMT26 Terminology Shared Task Track 1. At the official WMT26 evaluation, our system achieves 94.2% term success at 74.6 chrF++, with only two of the twenty-two submissions outperforming it on both metrics. On last year's benchmark, it also surpasses our GRPO-based system, despite being trained solely with ordinary supervised fine-tuning.