Translation with Thought: Difficulty-Adaptive Reasoning via Reinforcement Learning for Multi-Domain Machine Translation

Translation with Thought: Difficulty-Adaptive Reasoning via Reinforcement Learning for Multi-Domain Machine Translation Yongshi Ye1,3 , Biao Fu2,3, * , Chongxuan Huang2,3 , Yidong Chen2,3 , Xiaodong Shi1,2,3, * 1

Institute of Artificial Intelligence, Xiamen University 2 School of Informatics, Xiamen University 3 Key Laboratory of Digital Protection and Intelligent Processing of Intangible Cultural Heritage of Fujian and Taiwan (Xiamen University), Ministry of Culture and Tourism {yeyongshi,biaofu}@stu.xmu.edu.cn,mandel@xmu.edu.cn

arXiv:2607.29287v1 [cs.CL] 31 Jul 2026

Abstract Multi-domain machine translation (MDMT) poses a unique challenge due to varying levels of linguistic complexity across domains. Inspired by human translators’ ability to adapt reasoning effort based on difficulty, we propose TwT (Translation with Thought), a resourcerational framework that learns to modulate inference between intuitive and deliberate reasoning. TwT is trained in two stages: (1) supervised fine-tuning on difficulty-aware long chain-ofthought traces distilled from DeepSeek-R1 and rewritten by GPT-4o to reflect human-like reasoning economy, and (2) reinforcement learning with a hybrid reward to optimize translation quality and reasoning efficiency. Evaluated on 15 benchmarks spanning in-domain and outof-domain settings, as well as 3 seen and 59 unseen languages, with ablations across three backbone models, TwT-7B and TwT-14B outperform much larger SOTA reasoning models in translation quality, while reducing token usage by 32–60%. These results confirm that aligning translation behavior with cognitive principles enables robust generalization, high translation quality, and efficient reasoning in MDMT.

1

Introduction

Multi-domain machine translation (MDMT) remains a core challenge for language models due to significant variation in terminology, syntax, and style across domains. A key difficulty lies in the uneven distribution of complexity: some inputs are routine, while others require deeper reasoning to resolve ambiguity or domain-specific constructs. However, most MT systems translate all inputs uniformly, lacking mechanisms to adjust inference effort based on domain-specific complexity (Li et al., 2025; Liu et al., 2025). This contrasts with human translators, who adapt reasoning effort to input difficulty (Hvelplund, 2011; Gile and Lei, * Corresponding authors.

2020). They typically rely on fast, intuitive processing (System 1) for familiar content and slower, deliberate reasoning (System 2) when encountering Rich Points (Agar, 1994), such as ambiguous terminology, complex syntax, or cultural disparities. Because the density of such Rich Points varies across domains and correlates with translation difficulty (Lacruz, 2017), current MT systems still largely lack this adaptive reasoning ability. From the perspective of reasoning allocation, existing MT approaches fall into two extremes. On one end, standard large language model (LLM)based translators operate purely in System 1 mode: trained via supervised fine-tuning (SFT) on largescale parallel corpora (Xu et al., 2024a), they produce fluent outputs without explicit reasoning. While efficient, these models struggle with Rich Points and degrade in out-of-domain (OOD) or lowresource settings. Recent efforts have introduced Chain-of-Thought (CoT) prompting into translation (Wang et al., 2025a), but this does not fundamentally solve the problem, because the same reasoning pattern is applied uniformly regardless of input difficulty. Conversely, the emergence of large reasoning models (LRMs), such as DeepSeekR1 (Guo et al., 2025), represents a shift to the opposite extreme—an overcommitment to System 2. Reinforcement learning (RL) is often used to train these models to generate Long CoT traces, applying reasoning uniformly across inputs. This raises a natural question: Can RL serve as a bridge to align the model’s reasoning trajectory with the human translation process? To investigate this, we conduct two preliminary experiments (Section 3). The first examines Pure RL, where RL is applied directly to a base model without SFT on annotated reasoning traces. We find that the model rapidly collapses into repetitive, shallow templates, failing to develop domainspecific reasoning behaviors. The second explores RL with SFT, which fine-tunes on CoT traces be-

Figure 1: Case Study of Adaptive Thought. TwT switches between System 1 and System 2 based on complexity.

fore RL. While this setup produces longer reasoning, it lacks control over when such reasoning is needed, leading to verbose traces even for simple inputs. This indiscriminate reasoning may help reveal Rich Points, but often results in overthinking and excessive token usage, reducing efficiency and human alignment. Despite this, most reasoningbased MT methods still adopt either Pure RL (Feng et al., 2025a) or RL with SFT (Wang et al., 2025c), differing mainly in reward design. However, few attempt to align reasoning effort explicitly with input difficulty. Inspired by human cognitive flexibility, we propose TwT (Translation with Thought), a resourcerational framework for MDMT that learns to allocate inference effort based on input difficulty. As shown in Figure 1, TwT dynamically shifts its reasoning behavior according to input difficulty, using concise reasoning for routine inputs and deeper reasoning for domain-specific challenges. To implement this, TwT follows the RL with SFT pipeline. In the cold-start stage, it performs multi-agent distillation to construct difficulty-adaptive reasoning traces across domains: a domain-specialized teacher (DeepSeek-R1) generates diverse reasoning traces, and GPT-4o assesses input difficulty via Rich Points, rewriting the traces to match the appropriate reasoning depth. This process equips our student model with domain-sensitive reasoning and human-like inference modulation, supporting resource-rational translation. In the RL stage, we optimize the adaptive reasoning behavior seeded during cold-start by rewarding high-quality translations. Our hybrid reward combines translation quality metrics (BLEU and COMET) with a

repetition penalty, guiding the model toward efficient, domain-adaptive reasoning through outcomedriven learning. We conduct a comprehensive evaluation of TwT on 15 benchmarks across in-domain and OOD settings, as well as 3 seen and 59 unseen languages, and perform ablation studies on three different backbone models to assess generalization across both domain and linguistic axes. Our results demonstrate that TwT achieves performance competitive with or superior to SOTA LRMs (e.g., DeepSeek-R1, OpenAI-o1) and surpass strong MTspecialized baselines, while reducing token usage by 32–60%. Empirical analysis confirms that TwT effectively modulates reasoning effort according to task difficulty, leading to more coherent reasoning processes and more accurate translations. This validates the core intuition behind TwT: aligning reasoning effort with input difficulty yields both efficiency and quality gains.

2

Related Work

Recent MT studies increasingly explore explicit reasoning to improve translation quality, starting with shallow strategies such as disambiguation, domain recognition, and self-reflection (Chen et al., 2024; Feng et al., 2025b; Wang et al., 2024b; Hu et al., 2024). To support deeper reasoning, recent studies collect Long CoT traces via MCTS (Zhao et al., 2024) or multi-agent workflows (Wang et al., 2025a), then apply SFT. These traces emulate human translation workflows, improving both performance and interpretability (Chen et al., 2025; Liu et al., 2025). More recently, RL has emerged as a reasoning enhancer (Guo et al., 2025). Sev-

eral approaches optimize translation reasoning with verifiable rewards: R1-T1 uses COMET-based signals (He et al., 2025), MT-R1-Zero combines rulebased and neural metrics (Feng et al., 2025a), DeepTrans employs external LLMs (Wang et al., 2025b), and ExTrans adds exemplar-based guidance (Wang et al., 2025c). However, these methods overlook cognitive alignment; we model human-like reasoning to improve MDMT efficiency and quality.

3

Preliminary Analysis

3.1

Reasoning Collapse

Easy

Model

Hard

Quality Token Time Quality Token Time SFT-Parallel 64.85 General-CoT 61.54 Domain-CoT 67.36 SFT-Parallel 67.52 General-CoT 66.59 Domain-CoT 69.81

In-Domain 13 12 311 72 558 136

61.72 56.53 62.65

40 551 883

15 80 149

Out-of-Domain 13 7 59.11 354 96 60.55 634 125 62.52

36 517 840

8 97 138

Table 2: Performance by difficulty for SFT-Parallel, General-CoT, and Domain-CoT. Quality is computed as the average of BLEU, COMET, and COMETKIWI. Time denotes latency in milliseconds.

Reasoning Template

Rate

I will translate this Chinese sentence into English by identifying the key phrases and their meanings, and then constructing a coherent English sentence.

42.66%

I will translate this Chinese sentence into English by identifying the key phrases and their corresponding meanings.

24.56%

I will translate this Chinese sentence into English.

5.53%

Table 1: Top-3 reasoning template frequency.

We first investigate the R1-Zero paradigm (Pure RL); detailed training settings are given in Appendix E.1. This setup is motivated by recent findings that RL alone can induce spontaneous reasoning capabilities in math and code tasks (Guo et al., 2025). To test whether this emergence transfers to MDMT, we train models using GRPO with hybrid rewards. To rule out the possibility that KL regularization suppresses exploration (Yeo et al., 2025), we monitor token length dynamics both with and without the KL term. However, unlike the “Aha moments” observed in STEM tasks, our experiments reveal rapid mode collapse. As shown in Figures 2(a) and 2(b), reasoning traces quickly collapse into shallow patterns (≤ 100 tokens) regardless of the KL setting. Concretely, the model shifts toward high-frequency template recitation, suppressing diverse reasoning. As shown in Table 1, the top three templates account for about 73% of all generated traces in the Zh→En direction. This degeneration reveals a key misalignment: MDMT requires domain-aware reasoning, yet without proper initialization, the model produces reasoning that is too brief and overly templated to be elicited reliably. To address this, our cold-start phase explicitly initializes adaptive reasoning behavior before RL.

3.2

Reasoning Challenges

To evaluate the trade-off between reasoning depth and computational cost, we compare three Qwen2.5-7B-Instruct variants: SFT-Parallel (System 1), General-CoT (System 2), and Domain-CoT, which extends General-CoT with a domain-aware prompt. Details are given in Appendix E.2. Lack of Domain Awareness. As shown in Table 2, General-CoT performs poorly on in-domain data because its reasoning lacks explicit domain grounding, often defaulting to generic translations rather than domain-specific terminology and fixed expressions. By contrast, SFT-Parallel performs well in these cases by matching the distributional patterns of its training data. This same contrast also explains why General-CoT can be more competitive on OOD data, particularly on harder samples: when the input does not closely match the domain patterns seen in training, intermediate reasoning helps the model better handle syntax, ambiguity, and contextual inference than standard parallel SFT. Importantly, adding explicit domain reasoning largely restores in-domain quality, improving over General-CoT from 61.54 to 67.36 on Easy samples and from 56.53 to 62.65 on Hard samples. This confirms that lack of domain awareness is a major source of in-domain degradation. Reasoning Redundancy. However, domain awareness alone is not sufficient. General-CoT applies essentially the same reasoning strategy regardless of input difficulty, resulting in substantial redundancy. On Easy samples, this redundancy is clearly detrimental: despite using 23.9× more tokens in-domain (311 vs. 13) and 27.2× more on OOD data (354 vs. 13), it still underperforms SFT-

150 100

82

82

27

81

81

24

0

250

500

750

RL Training Steps

1000

18

1250

80 79

250

500

750

RL Training Steps

1000

1250

80 79

78 0

CometKiwi vs Training Step

83

30

21

50

COMET vs Training Step

83

COMET

200

BLEU

Response Length

250

BLEU vs Training Step

33

BLEU COMET CometKiwi BLEU+COMET BLEU+CometKiwi

CometKiwi

Response Length vs Training Step

300

0

250

500

750

RL Training Steps

1000

78

1250

0

250

500

750

RL Training Steps

1000

1250

(a) Pure RL training with KL regularization.

150 100

82

82

28

81

81

26

0

250

500

750

RL Training Steps

1000

1250

22

80 79

0

250

500

750

RL Training Steps

1000

1250

78

CometKiwi vs Step

83

30

24

50

COMET vs Training Step

83

COMET

200

BLEU

Response Length

250

BLEU vs Training Step

32

BLEU COMET CometKiwi BLEU+COMET BLEU+CometKiwi

CometKiwi

Response Length vs Training Step

300

80 79

0

250

500

750

RL Training Steps

1000

1250

78

0

250

500

750

RL Training Steps

1000

1250

(b) Pure RL training without KL regularization.

Figure 2: Training dynamics under pure RL using different quality rewards. While translation quality improves under all settings, pure RL training fails to induce extended translation reasoning traces.

Parallel by 3.31 and 0.93 quality points, respectively. Domain-CoT does not solve this problem. Although it restores in-domain quality, it further increases token usage to 42.9× on in-domain Easy samples (558 vs. 13) and 48.8× on OOD Easy samples (634 vs. 13). Similar patterns also hold on Hard samples. These results show that promptlevel domain awareness alone cannot resolve overthinking, motivating TwT, which adaptively shifts between System 1 and System 2 to jointly address domain sensitivity and reasoning efficiency.

4

Method

We propose TwT, a resource-rational approach that dynamically allocates reasoning effort by input difficulty, mimicking human translation process. As shown in Figure 3, training proceeds in two stages. 4.1

Cold Start

To align the model’s reasoning behavior with the human translation process, we construct a Difficulty-Adaptive CoT Dataset that equips the backbone LLM with adaptive reasoning capabilities. The dataset is curated through a multi-agent distillation-adaptation pipeline. We first prompt DeepSeek-R1 with domain-aware instructions to generate high-quality CoT traces tailored to different domains. These serve as the initial reasoning demonstrations. We then employ GPT-4o, which is verified to align best with human judgment (Ap-

pendix F.7), to assess input difficulty based on the theory of Rich Points. Specifically, difficulty is defined along four linguistic dimensions: sentence complexity, vocabulary rarity, grammatical divergence, and contextual nuance. The corresponding evaluation prompt is shown in Figures 10 and 11. Conditioned on this assessment, GPT-4o rewrites raw traces into adaptive formats: Easy inputs are reformulated into concise System 1 checks to minimize token usage, whereas Hard inputs retain comprehensive System 2 deliberations for structural and terminological verification. We then perform SFT on the backbone LLM using this compact dataset (∼7k examples). We define reasoning depth as the length of the generated reasoning trace (i.e., number of tokens). This process establishes an initial policy distribution over difficulty-aware reasoning strategies, enabling the model to autonomously modulate its reasoning depth. Representative examples are shown in Figures 13 and 14. The resulting data efficiency makes our approach particularly suitable for low-resource settings, demonstrating that robust adaptive reasoning can be achieved with a modest CoT-SFT seed dataset. 4.2

RL Training

To scale this adaptive reasoning behavior, we adopt the GRPO algorithm with hybrid quality rewards, which serve as outcome-driven constraints that encourage the model to identify Rich Points and al-

Stage1: Dataset Curation & Cold Start ① Domain-Aware CoT Data Curation

② Difficulty-Aware CoT Data Curation

③ CoT-FT

Translate the following {src_lang} text into {tgt_lang} while maintaining the domain style of the source text. Difficulty Evaluation

GPT-4o

Rewrite Traces

DeepSeek-R1

Model Size

Identify the domain & Generate domainspecific translation reasoning traces.

Estimate the translation difficulty & Optimize the reasoning trace to match its complexity level.

TwT-7B

TwT-14B

Stage2: RL Training Format Reward

Quality Reward

In-domain test set

TwTPenalty Repetition

TwT

In-Domain Test Set Comparison

������ = 1

������ =− 1

������ = ���� + �����

������ =− 2

� � = {�� = �� , …, �_{� + � − 1} } ������ = 1 −

|��� � � | |� � |

Figure 3: Overview of TwT training pipeline. TwT is first fine-tuned on difficulty-adaptive Long CoT traces distilled from DeepSeek-R1 and rewritten by GPT-4o for cognitive alignment. RL is then applied with a hybrid reward.

locate deep reasoning selectively—only where it leads to measurable quality improvements. By mitigating indiscriminate overthinking (Section 3.2), we improve reasoning efficiency and align the model with the resource-rational principle introduced in Section 1. The final reward r is crafted from three components to ensure alignment: r = rf + rq − λ · rrep Format Reward (rf ). We employ a binary format reward (rf ∈ {1, −1}) to strictly enforce the reasoning-translation structure defined in Figure 5. Hybrid Quality Reward (rq ). To prevent the model from producing plausible but functionally ineffective reasoning, we introduce a verifiable hybrid feedback signal. Building upon our preliminary analysis in Section 3.1, we identified a critical reward hacking phenomenon. As illustrated by the red curve in Figure 2(a), optimizing solely for a semantic metric (CometKiwi/COMET) leads to a significant metric divergence: despite high semantic reward scores, the lexical accuracy (BLEU) degrades rapidly during training. This confirms that the model hacks the reward by generating vague paraphrases or copying source tokens to maximize semantic similarity, effectively abandoning translation fidelity. To remedy this and strictly enforce alignment, we adopt a hybrid quality reward that combines BLEU (B) with COMET (C): ( B(ŷ, y) + C(x, ŷ, y) if rf = 1 rq = −2 if rf = −1

This design effectively stabilizes the optimization process. As evidenced by the training dynamics of our TwT models (Figure 8(a)), both the 7B and 14B models (Figures 8(b) and 8(c)) exhibit synchronous improvements in lexical and semantic metrics without divergence, validating that the hybrid signal successfully grounds the reasoning process in accurate translation outcomes. N-gram Repetition Penalty (rrep ). We penalize redundant reasoning to discourage degenerate loops and promote efficient token usage. For a given reasoning trace c, let G(c) denote the list of contiguous n-grams. In our experiments, we set n = 20. We compute the ratio of repeated n-grams to discourage repetitive, loop-like patterns: rrep = 1 −

|set (G(c))| ∈ [0, 1] |G(c)|

5

Experiments

5.1

Experimental Settings

Dataset. We use two datasets for training: (1) a curated set of 7K difficulty-adaptive Long CoT examples spanning 10 domains and three translation directions (De→En, En→Zh, Zh→En) for coldstart SFT, and (2) a separate 20K-sample dataset for RL, constructed from multi-domain parallel corpora. For evaluation, we adopt both in-domain test sets and diverse OOD benchmarks, and additionally include multilingual test sets covering both seen and unseen language pairs. Full dataset details are provided in Appendix B.

Laws

Method

Quality

News

Tokens

Quality

Science

Tokens

Quality

Subtitles

Tokens

DeepSeek-V3 Gemini-2.0-Flash GPT-4o

77.72 76.57 73.77

-

69.41 69.30 68.61

-

69.09 68.80 68.09

-

DeepSeek-R1 Gemini-2.0-Flash-Thinking OpenAI-o3-mini OpenAI-o1 GPT-5 QwQ-32B

77.72 76.21 71.61 73.85 76.28 71.84

577 702 428 478 784 667

68.47 68.24 68.18 68.67 69.07 67.88

498 1149 443 408 740 584

68.46 68.17 68.01 68.60 68.58 67.86

478 1092 385 367 606 563

SFT-Parallel-7B ALMA-7B-R ALMA-13B-R TowerInstruct-7B-v0.2 TowerInstruct-13B-v0.1 CoT-FT-7B MT-R1-Zero-7B SSR-X-Zero-7B mExTrans-7B

76.58 67.88 70.11 73.91 74.65 76.72 68.93 69.59 70.11

51 72 56 597

66.08 63.37 64.65 65.93 66.90 66.24 67.41 66.00 65.49

42 64 52 553

66.38 62.77 64.23 65.45 66.20 66.07 66.98 66.77 66.13

39 61 49 546

TwT-Qwen2.5-7B-Instruct TwT-Qwen2.5-14B-Instruct

75.35 76.58

310 320

68.42 68.62

311 285

68.24 68.32

294 272

Quality

Literary

Tokens

Quality

Large Language Models 62.93 56.83 62.96 57.41 62.92 57.47 Large Reasoning Models 61.81 514 53.95 61.98 708 57.23 62.42 355 56.97 62.62 340 57.28 62.74 519 56.27 61.80 584 55.09 MT-Specialized Models 62.87 55.96 59.38 54.52 60.22 55.30 61.07 55.32 61.89 56.09 62.76 29 55.57 61.97 55 55.80 61.89 39 55.54 59.52 476 54.40 Our Models 63.03 247 57.86 62.97 241 58.14

IT

Koran

Medical

Average

Tokens

Quality

Tokens

Quality

Tokens

Quality

Tokens

Quality

Tokens

-

66.97 66.54 66.37

-

57.73 58.13 57.78

-

69.16 70.21 69.34

-

66.23 66.24 65.54

-

574 781 546 521 859 863

66.28 66.20 65.99 66.45 66.57 61.44

593 345 343 403 492 583

57.48 58.13 56.42 57.94 58.71 55.81

790 677 511 506 751 963

69.01 69.73 68.40 69.20 70.14 66.68

667 415 346 441 531 735

65.40 65.74 64.75 65.58 66.05 63.55

586 734 420 433 660 693

52 69 54 610

68.02 64.20 64.54 66.78 67.18 67.82 65.60 61.22 60.50

35 56 36 452

58.05 55.04 55.71 50.05 49.93 57.57 55.52 55.53 55.41

45 71 46 604

70.21 67.46 68.40 70.73 71.49 70.38 64.43 63.99 63.00

46 71 50 565

65.52 61.83 62.89 63.66 64.29 65.39 63.33 62.56 61.82

42 65 48 551

281 354

68.07 68.37

222 234

58.92 59.35

269 336

70.29 70.59

262 287

66.27 66.62

274 291

Table 3: In-domain translation performance across eight domains, averaged over En→Zh, Zh→En, and De→En. The Bold and underlined values denote the highest and second highest scores, respectively. Method

Conversation Quality

Tokens

Ecommerce Quality

Social

Tokens

DeepSeek-V3 Gemini-2.0-Flash GPT-4o

68.42 68.77 68.75

-

66.34 66.26 66.50

-

DeepSeek-R1 Gemini-2.0-Flash-Thinking OpenAI-o3-mini OpenAI-o1 GPT-5

67.06 68.36 68.04 68.47 68.40

534 1204 290 327 448

64.45 65.83 65.87 65.80 65.49

552 822 363 399 609

SFT-Parallel-7B ALMA-7B-R ALMA-13B-R CoT-FT-7B MT-R1-Zero-7B SSR-X-Zero-7B mExTrans-7B

65.64 64.71 66.03 65.43 66.59 65.70 63.47

31 53 37 464

63.33 62.54 63.37 63.26 64.16 63.46 61.74

45 66 50 566

TwT-Qwen2.5-7B-Instruct TwT-Qwen2.5-14B-Instruct

67.72 67.77

231 240

65.71 65.79

273 309

Quality

Culture Tokens

Large Language Models 66.10 66.10 66.10 Large Reasoning Models 64.11 554 65.50 1081 65.58 372 65.38 405 65.14 652 MT-Specialized Models 62.54 62.68 63.51 62.06 42 63.69 65 63.31 49 61.20 555 Our Models 65.48 269 65.47 298

CommonSense

Average

Quality

Tokens

Quality

Tokens

Quality

Tokens

69.65 69.02 69.01

-

65.96 65.23 65.89

-

67.29 67.08 67.25

-

68.16 68.42 67.55 67.96 68.33

560 1220 596 542 984

64.34 65.44 64.45 64.78 64.09

602 2335 436 392 530

65.62 66.71 66.30 66.48 66.29

561 1332 411 413 645

65.34 66.63 60.94 64.66 66.23 64.35 65.11

54 79 66 631

61.03 62.02 62.91 61.08 62.32 62.18 59.92

33 51 34 470

63.58 63.71 63.35 63.30 64.60 63.80 62.29

41 63 47 537

67.82 68.55

352 333

64.48 64.72

219 259

66.25 66.46

269 288

Table 4: OOD translation performance across five domains, averaged over En→Zh, Zh→En, and De→En.

Implementation Details. For cold start, we use LLaMA-Factory1 (Zheng et al., 2024) with Qwen2.5-7B-Instruct, Qwen2.5-14B-Instruct, and Gemma-2-9B-IT as backbones. We train them on the 7K difficulty-adaptive Long CoT examples for 1 epoch with full-parameter optimization on 8 NVIDIA A100 80GB GPUs, using AdamW with a learning rate of 1e−5, a total batch size of 32, a cosine learning rate scheduler, a warm-up ratio of 0.1, a maximum input sequence length of 4096, and DeepSpeed ZeRO Stage 3. The coldstart stage completes within 10 minutes. For RL, we use verl2 (Sheng et al., 2025) and train for 1 epoch on 8 NVIDIA A100 80GB GPUs with a total batch size of 16, rollout number 16, rollout temperature 1.0, learning rate 1e−6, KL loss coefficient β = 1e−3, maximum response length 2048, and repetition penalty with n = 20. RL training 1 2

https://github.com/hiyouga/LLaMA-Factory https://github.com/volcengine/verl

takes about 10 hours. During inference, we use vLLM3 (Kwon et al., 2023) for efficient decoding with temperature 0.0 and repetition penalty 1.05. Metrics. We report Quality, defined as the average of BLEU, COMET (Rei et al., 2020), and CometKiwi (Rei et al., 2022), and Tokens, the average length of the generated CoT. The full metric breakdowns are provided in Appendix I. Baselines. We compare our method against three categories of models: (1) general-purpose LLMs such as DeepSeek-V3 (DeepSeek-AI et al., 2024), Gemini-2.0-Flash (DeepMind, 2024), GPT4o (OpenAI et al., 2024), and open-source models like LLaMA3.1-8B-Instruct (Grattafiori et al., 2024), Gemma-2-9B-IT (Gemma Team et al., 2024), and Qwen2.5 series (7B, 14B, 32B) (Yang et al., 2024). (2) reasoning-oriented LRMs, such as DeepSeek-R1 (Guo et al., 2025), Gemini3

https://github.com/vllm-project/vllm

En→Zh

Method

Quality

Zh→En

Tokens

Quality

Qwen2.5-7B-Instruct Gemma-2-9B-IT

67.87 66.50

-

50.91 52.64

ALMA-7B-R Tower-Plus-9B SFT-Parallel-7B MT-R1-Zero-7B SSR-X-Zero-7B mExTrans-7B

64.08 69.85 69.44 67.52 68.55 66.70

62 50 537

54.52 57.41 55.96 55.80 56.52 54.40

TwT-Qwen2.5-7B-Instruct TwT-Gemma-2-9B-IT

69.99 69.07

298 227

57.86 58.12

De→En

Tokens

Quality

Tokens

Large Language Models 62.90 60.64 MT-Specialized Models 62.89 67.04 66.15 69 62.56 66 54 62.86 45 610 60.74 544 Our Models 281 66.45 256 249 66.95 218

En→X

X→En

Average

Quality

Tokens

Quality

Tokens

Quality

Tokens

37.61 54.28

-

56.99 66.38

-

55.25 60.09

-

44.68 46.46 32.44 40.98 40.43 42.72

358 306 1047

44.71 63.02 55.70 58.33 58.19 55.99

74 48 731

54.18 60.76 55.94 57.04 57.31 56.11

126 101 694

41.13 54.65

483 280

58.84 67.04

328 257

58.85 63.17

329 246

Table 5: Results on seen and unseen language directions. En, Zh, and De are seen languages, while X denotes unseen languages; En→X and X→En report averages over English↔unseen-language directions.

Method TwT-Qwen2.5-7B-Instruct w/o RP w/o RP + w/o Adaptive CoT w/o RP + w/o Cold Start SFT only w/ Adaptive CoT SFT only w/ domain-aware CoT SFT only w/ general CoT

In-Domain

Out-of-Domain

Quality

Tokens

Quality

Tokens

64.77 64.52 64.68 64.18 62.47 61.65 61.39

278 301 748 62 253 479 531

66.28 66.12 66.17 65.29 64.09 63.47 63.11

263 271 720 59 219 452 519

64.71 63.33 63.68

-

66.28 64.44 66.05

263 298 220

Backbone Models 61.99 59.84 59.93 Our Models TwT-Qwen2.5-7B-Instruct 64.77 278 TwT-Llama-3.1-8B-Instruct 63.20 305 TwT-Gemma-2-9B-IT 64.71 231

Qwen2.5-7B-Instruct Llama-3.1-8B-Instruct Gemma-2-9B-IT

Table 6: Ablation study on in-domain and OOD translation test sets. Results are averaged at the dataset level for each setting. RP = repetition penalty.

2.0-Flash-Thinking (DeepMind, 2025), OpenAI o1 (Jaech et al., 2024), o3-mini (OpenAI, 2025b), GPT-5 (OpenAI, 2025a), and QwQ-32B (QwenTeam, 2025); (3) MT-specialized models, including non-reasoning LLMs such as TowerInstruct (Alves et al., 2024), ALMA-R (Xu et al., 2024a,b), SFTParallel (Qwen2.5-7B-Instruct fine-tuned on 27K parallel pairs), CoT-FT (Hu et al., 2024), and Tower-Plus-9B (Rei et al., 2025); with reasoningoriented models including MT-R1-Zero (Feng et al., 2025a), mExTrans (Wang et al., 2025c), and SSRX-Zero (Yang et al., 2025). For fairer comparison, Appendix C details the training data size of MTspecialized baselines. 5.2

Cross-Domain Generalization

In-Domain. As shown in Table 3, TwT-14B achieves a SOTA average quality score of 66.62, outperforming both strong open-source LRMs such

as DeepSeek-R1 (65.40) and closed-source models including GPT-5 (66.05). Notably, our smaller variant, TwT-7B, also surpasses dedicated MT systems, demonstrating that our method scales effectively with model size. Compared to mExTrans7B, which is also trained under the R1 paradigm, TwT-7B yields a substantial improvement of +4.45 points while reducing reasoning overhead by 50.27%, highlighting the superior efficiency of our difficulty-adaptive mechanism. The benefits of adaptive reasoning are especially evident in structurally complex and culturally nuanced domains. In the Literary domain, TwT-14B consistently outperforms three representative paradigms: it exceeds the System 1 baseline SFT-Parallel by +2.18 points, the heavy-reasoning System 2 model DeepSeek-R1 by +4.19 points, and the pure RL-based MT-R1Zero-7B by +2.34 points. These results suggest that, for multi-domain translation, neither shallow System 1 execution, indiscriminate System 2 overthinking, nor unguided pure RL exploration alone yields optimal performance. In contrast, our model autonomously modulates reasoning depth to strike a more effective balance between literal accuracy and stylistic adequacy. OOD. TwT-7B achieves a strong average score of 66.25 on five OOD test sets, surpassing MTspecialized baselines such as SFT-Parallel-7B (63.58), which exhibit poor generalization under domain shift. Moreover, TwT-7B exhibits a favorable quality-efficiency trade-off, surpassing DeepSeek-R1 while reducing reasoning overhead by 292 tokens. Even in unfamiliar domains, it avoids excessive deliberation by leveraging compact, internalized translation procedures. These results indicate that TwT captures domain-agnostic

translation logic rather than relying on domainspecific memorization, approaching the SOTA performance of GPT-4o at a fraction of the computational cost and model size. 5.3

Multilingual Generalization

On seen directions, TwT-7B improves Zh→En performance by +6.95 over its base model and +1.90 over SFT-Parallel-7B, under the same training data. On unseen directions (En→X)4 , TwT-Gemma-2-9B-IT, trained on only 27K examples, outperforms the multilingual Tower-Plus-9B, built on the same backbone but trained on 286K examples, by a substantial margin of +8.19. Overall, TwT achieves the highest average score across all directions, indicating that its reasoning mechanism generalizes beyond language boundaries and captures transferable alignment strategies. 5.4

Ablation Study

We conduct a comprehensive ablation study to assess the contribution of each component in the TwT framework and validate its generalizability across backbone architectures (Table 6). Removing the repetition penalty (w/o RP) leads to a slight quality drop and longer outputs, indicating its role as a regularizer rather than a performance driver. In contrast, removing both the repetition penalty and the difficulty-adaptive rewriting (w/o RP + w/o Adaptive CoT) results in comparable quality but significantly increases reasoning length (from 278 to 748 tokens), highlighting the critical role of adaptive rewriting in controlling verbosity and ensuring inference efficiency. We further examine the necessity of the two-stage training pipeline. Eliminating the cold-start SFT phase (w/o RP + w/o Cold Start) causes performance degradation and length collapse (to 62 tokens), suggesting that RL alone fails to induce structured reasoning behavior. To isolate the impact of SFT data quality, we compare three variants: difficulty-adaptive CoT yields the best result (62.47), followed by domain-aware (61.65) and general CoT (61.39), showing a clear performance hierarchy. Still, only the full TwT pipeline achieves the highest score (64.77), confirming that RL is indispensable for turning the adaptive reasoning patterns from mere imitation into an internalized and optimized translation strategy. Lastly, we apply TwT to three backbone models: Qwen2.5-7B, 4

En→X and X→En are averaged over 59 unseen languages from FLORES+; see Appendix B.6 for the full list.

Llama-3.1-8B, and Gemma-2-9B, and observe consistent improvements. For instance, TwT improves Gemma-2-9B-IT from 59.93 to 64.71 in-domain and from 63.68 to 66.05 OOD, demonstrating that TwT is a model-agnostic framework that robustly enhances translation reasoning regardless of the underlying architecture.

6

Empirical Analysis

6.1

Human Reasoning Alignment

To benchmark TwT’s reasoning against human cognition, we conduct a qualitative analysis on 10 Zh→En examples, with expert commentary from a translation studies faculty member. A representative case is shown in Appendix G.2. Cognitive Convergence. The analysis revealed that TwT’s reasoning exhibits strong parallels with human translators in early-stage decision-making. Specifically, TwT effectively (1) identifies translation domain and stylistic register, (2) handles complex sentence structures with appropriate syntactic parsing, and (3) demonstrates context-aware terminology adaptation. For example, it consistently distinguishes between literary and technical expressions and adjusts lexical choices accordingly. Its structured CoT mirrors key aspects of professional reasoning—such as coherence maintenance, discourse flow control, and sensitivity to stylistic norms—indicating that TwT has internalized domain-aware reasoning behavior resembling human translation logic. Pragmatic Divergence. Despite these strengths, TwT still shows gaps compared with expert translators. It occasionally struggles with cross-sentence consistency in terminology, especially when handling long-form repetitions or abbreviated references. Moreover, its output lacks fine-grained control over tone, idiomaticity, and cultural adaptation, which human translators adjust based on pragmatic context and target audience. These issues suggest that TwT’s reasoning remains less flexible in discourse-level adaptation, reflecting the absence of high-level pragmatic awareness. Future work seeks to bridge these gaps by integrating processoriented feedback, thereby fostering deeper pragmatic alignment with human cognitive processes. 6.2

Reasoning Redundancy Reduction

To assess whether TwT eliminates unnecessary computation, we employ DeepSeek-V3.2 to detect six

Over-segmentation Unnecessary linguistic explanation Semantic repetition Irrelevant information Redundant alternative translations Low-density long descriptions

Semantic Space Exploration (PCA)

Not Resolved (%)

87.3 93.2 96.9 93.7 96.5 97.2

12.7 6.8 3.1 6.3 3.5 2.8

Pairwise Reasoning Similarity Distribution

Pure-RL TwT (Ours)

0.4

3.0

0.2

2.5

0.0

2.0

Density

Resolved (%)

Principal Component 2

Redundancy Type

Easy

Method DeepSeek-R1 Gemini-2.0-Flash-Thinking OpenAI-o3-mini OpenAI-o1 QwQ-32B General-CoT TwT-Qwen2.5-7B-Instruct TwT-Qwen2.5-14B-Instruct

Medium

Hard

All

Quality

Tokens

Quality

Tokens

Quality

Tokens

Quality

Tokens

66.30 66.62 66.57 67.14 62.91 61.54 67.93 67.98

480 432 280 302 577 311 216 207

63.82 64.34 63.44 64.26 61.92 60.27 64.85 65.23

556 754 430 432 758 433 272 300

62.15 62.93 61.44 62.49 60.52 56.53 62.78 63.26

579 1035 596 578 844 551 332 378

64.09 64.63 63.82 64.63 61.78 59.45 65.19 65.49

538 740 435 437 726 432 273 295

Table 8: In-domain performance by difficulty level.

distinct forms of reasoning redundancy across 15 domains, utilizing the prompt provided in Figure 12. As shown in Table 7, TwT demonstrates exceptional efficiency, successfully resolving over 94% of the redundant steps observed in the SFTwith-RL baseline described in Section 3.2. Notably, it achieves a 97.2% resolution rate for Low-Density Long Descriptions, verifying its ability to compress verbose reasoning into high-density insights, while maintaining 87.3% resolution for structural issues like Over-Segmentation. These results confirm that the model has internalized resource-rational adaptive reasoning behavior, effectively activating System 2 reasoning for Rich Points while avoiding unnecessary elaboration on straightforward segments. 6.3

Translation Difficulty Adaptation

To better understand how models adapt their reasoning behavior to translation difficulty, we group the in-domain test set into three difficulty levels (Easy, Medium, Hard) estimated by DeepSeek-V3 (prompt in Figure 11). Table 8 reports the average quality and the response length for each group, averaged across all domains. Results indicate that TwT effectively addresses the reasoning redundancy of General-CoT through resource-rational allocation. On Easy inputs, it reduces token usage by 33% while improving quality by +6.4 points; conversely, on Hard inputs, it focuses on performance, achieving a substantial +6.7 points quality gain. Moreover, TwT models consistently outperform all SOTA LRMs across all difficulty levels in terms of quality, while maintaining significantly shorter reasoning traces—reducing average token usage by 32% compared to OpenAI-o3-mini and by 60% compared to Gemini-2.0-Flash-Thinking. This highlights TwT’s ability to generate concise, difficulty-aware reason-

1.5

0.2 1.0 0.4

Table 7: Resolution rates of six redundancies by TwT.

Pure-RL TwT (Ours)

0.5

0.6

0.4

0.2

0.0

Principal Component 1

0.2

0.4

0.0 0.0

0.2

0.4

0.6

Cosine Similarity

0.8

1.0

Figure 4: CoT trace similarity comparison. Pure RL vs TwT in space (Left) and distribution (Right).

ing while reducing overthinking. 6.4

Reasoning Collapse Mitigation

In Section 3.1, we identified a critical failure mode of pure RL, where reasoning rapidly degenerates into shallow, repetitive templates. To verify whether TwT successfully mitigates this reasoning collapse, we conduct a semantic diversity analysis across 15 domains. We compute pairwise cosine similarities of generated reasoning traces using multilingual Sentence-BERT5 . Pure RL exhibits severe redundancy with a mean similarity of 0.89, whereas TwT significantly reduces this metric to 0.51. This divergence is visually corroborated by Figure 4, where the PCA projection (Left) shows pure RL confined to tight, isolated clusters compared to the broad semantic manifold of TwT, and the similarity histogram (Right) confirms that TwT diffuses the sharp redundancy peak of the baseline into a balanced distribution. These results show that TwT overcomes template dependency and encourages genuine reasoning. 6.5

Further Analysis

Appendix F provides additional analyses of TwT, including MQM error types, KL ablation, training dynamics, domain-aware prompting, inference cost, language consistency, and failure cases.

7

Conclusion

In this work, we present TwT, a resource-rational translation model that adapts reasoning effort to input difficulty. TwT combines difficulty-aware SFT and hybrid-reward RL to balance System 1 and System 2 behavior. Evaluated across diverse domains and languages, TwT matches or surpasses SOTA LRMs while reducing token usage by 32–60%, validating the effectiveness of aligning translation with human reasoning economy. 5

sentence-transformers/paraphrase-multilingualMiniLM-L12-v2

Limitations While TwT achieves robust performance across multiple domains, several limitations remain. First, the RL training data is randomly sampled without controlling for difficulty distribution, which may result in an imbalanced mix of easy, medium, and hard inputs. Second, the reasoning traces distilled from proprietary LLMs (e.g., DeepSeek-R1) may carry over implicit biases or domain preferences inherent in those models. Although our current setup yields consistent improvements, such biases could influence the reasoning behavior or stylistic tendencies of TwT. Finally, our current reward design does not incorporate difficulty-aware reward shaping. In particular, no length-based reward is applied to encourage concise reasoning on simple inputs and more detailed analysis for complex ones. Incorporating such adaptive rewards may further enhance the model’s ability to adjust reasoning depth based on input complexity in MDMT. We leave this direction for future work.

Acknowledgment This work is supported by the National Science and Technology Major Project (Grant No. 2022ZD0116101), the National Natural Science Foundation of China (NSFC) under Grant No. 62206295, the Major Scientific Research Project of the State Language Commission in the 13th FiveYear Plan (Grant No. WT135-38), the public technology service platform project of Xiamen City (No. 3502Z20231043). In addition, we used a large language model to assist in polishing the visualizations in Figure 1 and generating certain decorative visual elements in Figure 3.

References M. Agar. 1994. Language Shock: Understanding The Culture Of Conversation. HarperCollins. Roee Aharoni and Yoav Goldberg. 2020. Unsupervised domain clusters in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7747– 7763, Online. Association for Computational Linguistics. Duarte M. Alves, José Pombal, Nuno M. Guerreiro, Pedro H. Martins, João Alves, Amin Farajian, Ben Peters, Ricardo Rei, Patrick Fernandes, Sweta Agrawal, Pierre Colombo, José G. C. de Souza, and André F. T. Martins. 2024. Tower: An open multilingual large language model for translation-related tasks. Preprint, arXiv:2402.17733.

Andong Chen, Lianzhang Lou, Kehai Chen, Xuefeng Bai, Yang Xiang, Muyun Yang, Tiejun Zhao, and Min Zhang. 2024. DUAL-REFLECT: Enhancing large language models for reflective translation through dual learning feedback mechanisms. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 693–704, Bangkok, Thailand. Association for Computational Linguistics. Andong Chen, Yuchen Song, Wenxin Zhu, Kehai Chen, Muyun Yang, Tiejun Zhao, and Min zhang. 2025. Evaluating o1-like llms: Unlocking reasoning for translation through comprehensive analysis. Preprint, arXiv:2502.11544. Google DeepMind. 2024. Introducing gemini 2.0: our new ai model for the agentic era. https: //blog.google/technology/google-deepmind/ google-gemini-ai-update-december-2024/ #ceo-message. Accessed: 2025-04-21. Google DeepMind. 2025. Gemini 2.0 flash thinking. https://ai.google.dev/gemini-api/ docs/changelog. Accessed: 2025-04-21. DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, and 181 others. 2024. Deepseek-v3 technical report. Preprint, arXiv:2412.19437. Zhaopeng Feng, Shaosheng Cao, Jiahan Ren, Jiayuan Su, Ruizhe Chen, Yan Zhang, Zhe Xu, Yao Hu, Jian Wu, and Zuozhu Liu. 2025a. Mt-r1-zero: Advancing llm-based machine translation via r1-zero-like reinforcement learning. Preprint, arXiv:2504.10160. Zhaopeng Feng, Yan Zhang, Hao Li, Bei Wu, Jiayu Liao, Wenqiang Liu, Jun Lang, Yang Feng, Jian Wu, and Zuozhu Liu. 2025b. TEaR: Improving LLM-based machine translation with systematic self-refinement. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 3922–3938, Albuquerque, New Mexico. Association for Computational Linguistics. Gemma Team, Morgane Riviere, and 1 others. 2024. Gemma 2: Improving open language models at a practical size. Preprint, arXiv:2408.00118. Daniel Gile and Victoria Lei. 2020. Translation, effort and cognition. In The Routledge handbook of translation and cognition, pages 263–278. Routledge. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad AlDahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, and 180 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. Preprint, arXiv:2501.12948. Jie He, Tao Wang, Deyi Xiong, and Qun Liu. 2020. The box is in the pen: Evaluating commonsense reasoning in neural machine translation. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3662–3672, Online. Association for Computational Linguistics.

Isabel Lacruz. 2017. Cognitive effort in translation, editing, and post-editing. The handbook of translation and cognition, pages 386–401. Zihao Li, Shaoxiong Ji, and Jörg Tiedemann. 2025. Testtime scaling of reasoning models for machine translation. Preprint, arXiv:2510.06471. Sinuo Liu, Chenyang Lyu, Minghao Wu, Longyue Wang, Weihua Luo, Kaifu Zhang, and Zifu Shang. 2025. New trends for modern machine translation with large reasoning models. Preprint, arXiv:2503.10351.

Minggui He, Yilun Liu, Shimin Tao, Yuanchang Luo, Hongyong Zeng, Chang Su, Li Zhang, Hongxia Ma, Daimeng Wei, Weibin Meng, Hao Yang, Boxing Chen, and Osamu Yoshie. 2025. R1-t1: Fully incentivizing translation capability in llms via reasoning learning. Preprint, arXiv:2502.19735.

NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, and 20 others. 2024. Scaling neural machine translation to 200 languages. Nature, 630(8018):841–846.

Tianxiang Hu, Pei Zhang, Baosong Yang, Jun Xie, Derek F. Wong, and Rui Wang. 2024. Large language model for multi-domain translation: Benchmarking and domain CoT fine-tuning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5726–5746, Miami, Florida, USA. Association for Computational Linguistics.

OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Madry, ˛ Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, and 401 others. 2024. Gpt-4o system card. Preprint, arXiv:2410.21276.

Kristian Tangsgaard Hvelplund. 2011. Allocation of cognitive resources in translation: An eye-tracking and key-logging study. Frederiksberg: Copenhagen Business School (CBS).

OpenAI. 2025a. Introducing GPT-5. https://openai. com/zh-Hans-CN/index/introducing-gpt-5/.

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, Alex Iftimie, Alex Karpenko, Alex Tachard Passos, Alexander Neitz, Alexander Prokofiev, Alexander Wei, Allison Tam, Ally Bennett, Ananya Kumar, and 242 others. 2024. Openai o1 system card. Preprint, arXiv:2412.16720. Tom Kocmi, Rachel Bawden, Ondřej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Thamme Gowda, Yvette Graham, Roman Grundkiewicz, Barry Haddow, Rebecca Knowles, Philipp Koehn, Christof Monz, Makoto Morishita, Masaaki Nagata, Toshiaki Nakazawa, Michal Novák, Martin Popel, and Maja Popović. 2022. Findings of the 2022 conference on machine translation (WMT22). In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1–45, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, SOSP ’23, page 611–626, New York, NY, USA. Association for Computing Machinery.

OpenAI. 2025b. Openai o3-mini. https://openai. com/index/openai-o3-mini/. Accessed: 202504-21. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. Qwen-Team. 2025. Qwq-32b: Embracing the power of reinforcement learning. Ricardo Rei, Nuno M. Guerreiro, José Pombal, João Alves, Pedro Teixeirinha, Amin Farajian, and André F. T. Martins. 2025. Tower+: Bridging generality and translation specialization in multilingual llms. Preprint, arXiv:2506.17080. Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685–2702, Online. Association for Computational Linguistics. Ricardo Rei, Marcos Treviso, Nuno M. Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José G. C. de Souza, Taisiya Glushkova, Duarte Alves, Luisa Coheur, Alon Lavie, and André F. T.

Martins. 2022. CometKiwi: IST-unbabel 2022 submission for the quality estimation shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 634–645, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. Preprint, arXiv:1707.06347. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300. Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2025. Hybridflow: A flexible and efficient rlhf framework. In Proceedings of the Twentieth European Conference on Computer Systems, EuroSys ’25, page 1279–1297, New York, NY, USA. Association for Computing Machinery. Liang Tian, Derek F. Wong, Lidia S. Chao, Paulo Quaresma, Francisco Oliveira, Yi Lu, Shuo Li, Yiming Wang, and Longyue Wang. 2014. UM-corpus: A large English-Chinese parallel corpus for statistical machine translation. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC‘14), pages 1837–1842, Reykjavik, Iceland. European Language Resources Association (ELRA). Jiaan Wang, Fandong Meng, Yunlong Liang, and Jie Zhou. 2025a. Drt: Deep reasoning translation via long chain-of-thought. Preprint, arXiv:2412.17498. Jiaan Wang, Fandong Meng, and Jie Zhou. 2025b. Deep reasoning translation via reinforcement learning. Preprint, arXiv:2504.10187. Jiaan Wang, Fandong Meng, and Jie Zhou. 2025c. Extrans: Multilingual deep reasoning translation via exemplar-enhanced reinforcement learning. Preprint, arXiv:2505.12996. Longyue Wang, Siyou Liu, Chenyang Lyu, Wenxiang Jiao, Xing Wang, Jiahao Xu, Zhaopeng Tu, Yan Gu, Weiyu Chen, Minghao Wu, Liting Zhou, Philipp Koehn, Andy Way, and Yulin Yuan. 2024a. Findings of the WMT 2024 shared task on discourse-level literary translation. In Proceedings of the Ninth Conference on Machine Translation, pages 699–700, Miami, Florida, USA. Association for Computational Linguistics. Longyue Wang, Zhaopeng Tu, Yan Gu, Siyou Liu, Dian Yu, Qingsong Ma, Chenyang Lyu, Liting Zhou, ChaoHong Liu, Yufeng Ma, Weiyu Chen, Yvette Graham, Bonnie Webber, Philipp Koehn, Andy Way, Yulin Yuan, and Shuming Shi. 2023. Findings of the WMT

2023 shared task on discourse-level literary translation: A fresh orb in the cosmos of LLMs. In Proceedings of the Eighth Conference on Machine Translation, pages 55–67, Singapore. Association for Computational Linguistics. Yutong Wang, Jiali Zeng, Xuebo Liu, Fandong Meng, Jie Zhou, and Min Zhang. 2024b. TasTe: Teaching large language models to translate through selfreflection. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6144–6158, Bangkok, Thailand. Association for Computational Linguistics. Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. 2024a. A paradigm shift in machine translation: Boosting translation performance of large language models. In The Twelfth International Conference on Learning Representations. Haoran Xu, Amr Sharaf, Yunmo Chen, Weiting Tan, Lingfeng Shen, Benjamin Van Durme, Kenton Murray, and Young Jin Kim. 2024b. Contrastive preference optimization: Pushing the boundaries of LLM performance in machine translation. In Forty-first International Conference on Machine Learning. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, and 23 others. 2024. Qwen2.5 technical report. Preprint, arXiv:2412.15115. Wenjie Yang, Mao Zheng, Mingyang Song, Zheng Li, and Sitong Wang. 2025. Ssr-zero: Simple selfrewarding reinforcement learning for machine translation. Preprint, arXiv:2505.16637. Binwei Yao, Ming Jiang, Tara Bobinac, Diyi Yang, and Junjie Hu. 2024. Benchmarking machine translation with cultural awareness. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13078–13096, Miami, Florida, USA. Association for Computational Linguistics. Edward Yeo, Yuxuan Tong, Morry Niu, Graham Neubig, and Xiang Yue. 2025. Demystifying long chain-of-thought reasoning in llms. Preprint, arXiv:2502.03373. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, and 16 others. 2025. Dapo: An open-source llm reinforcement learning system at scale. Preprint, arXiv:2503.14476. Yu Zhao, Huifeng Yin, Bo Zeng, Hao Wang, Tianqi Shi, Chenyang Lyu, Longyue Wang, Weihua Luo, and Kaifu Zhang. 2024. Marco-o1: Towards open reasoning models for open-ended solutions. Preprint, arXiv:2411.14405.

Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. 2024. LlamaFactory: Unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 400–410, Bangkok, Thailand. Association for Computational Linguistics.

A

GRPO Algorithm

GRPO (Shao et al., 2024) extends PPO (Schulman et al., 2017) by removing the dependency on a value model and instead leveraging group-wise relative rewards estimation among sampled responses for more stable and efficient policy updates. Given a query x, the model samples a group of G responses {yi }G i=1 , each scored with reward ri . The normalized advantage for each sample is computed as: Ai =

ri − mean({r}G j=1 ) std({r}G j=1 )

i=1

old

A conversation between User and Assistant. The user asks a translation question, and the Assistant solves it. The Assistant first thinks about the translation reasoning process in the mind, and then provides the final translation. The translation reasoning process and the final translation are enclosed within <think> </think> and <answer> </answer> tags, respectively, i.e., <think> translation reasoning process here </think> <answer> final translation here </answer>.\n\n User: {Translation question}.\n Assistant: <think> Figure 5: Template for pure RL in MT task.

.

(1)

Then GRPO optimizes the policy model πθ by maximizing the following objective: JGRPO (θ) = Ex∼D, {yi }G ∼πθ

Template

German-English multi-domain • The dataset (Aharoni and Goldberg, 2020), including five distinct domains: IT, Law, Medical, Koran, and Subtitles.

(·|x)

G  π (y | x) 1 X θ i min Ai , G πθold (yi | x) i=1  π (y | x)   θ i clip , 1 − ϵ, 1 + ϵ Ai πθold (yi | x) !
− βDKL πθ πref ,

(2) where πθold and πθ are the old and current policies, ϵ is the PPO clipping threshold, and β controls the weight of the KL regularization.

• The English-Chinese UM-Corpus (Tian et al., 2014), covering four domains: News, Laws, Subtitles, and Science. • The Chinese-English GuoFeng-Webnovel dataset (Wang et al., 2023, 2024a) from WMT23 and WMT24 literary translation tasks, representing the Literary domain.

B

Datasets

For each domain, we randomly select 2K sentence pairs with a minimum source sentence length of 20 words (characters for Chinese) to ensure meaningful reasoning potential. This results in a total of 20K training samples used for RL training.

B.1

Details of Cold Start Data

B.3

For cold-start SFT, we collect a curated dataset of about 7K difficulty-adaptive Long CoT examples spanning 10 domains and three major translation directions: De→En, En→Zh, and Zh→En. The data is constructed via domain-aware generation with DeepSeek-R1 followed by difficulty-adaptive rewriting with GPT-4o (see Section 4.1). Detailed dataset statistics are presented in Figure 6. B.2

Details of RL Training Data

We collect a diverse MDMT dataset for RL training across languages and domains. Specifically, we sample from the following sources:

In-Domain Test Data

For in-domain evaluation, we use the official test sets associated with the corpora in Appendix B.2. For the Literary domain, we merge the valid_1, valid_2, test_1, and test_2 subsets to form a comprehensive test set. The data statistics is illustrated in Table 9. B.4

Out-of-Domain Test Data

For out-of-domain evaluation, we consider a diverse set of test sets spanning multiple language pairs and domains. Specifically, the Conversation, Ecommerce, and Social domains are drawn from the WMT22 shared tasks (Kocmi et al.,

0XOWLGRPDLQ'H(Q0HGLFDO

0XOWLGRPDLQ'H(Q6XEWLWOHV

0XOWLGRPDLQ'H(Q/DZ 



*XR)HQJ:HEQRYHO/LWHUDU\



 80FRUSXV(Q=K/DZV



0XOWLGRPDLQ'H(Q.RUDQ

 

 



0XOWLGRPDLQ'H(Q,7

80FRUSXV(Q=K1HZV 80FRUSXV(Q=K6XEWLWOHV

80FRUSXV(Q=K6FLHQFH

Figure 6: Distribution of our curated difficulty-adaptive Long CoT data.

Domain

Num

Domain

Num

En→Zh Laws En→Zh Subtitles En→Zh Science En→Zh News Zh→En Literary

456 597 503 1500 3038

De→En IT De→En Koran De→En Medical De→En Law De→En Subtitles

2000 2000 2000 2000 2000

Table 9: In-domain test sets and the number of samples for En↔Zh and De→En translation tasks. Domain

Num

Domain

Num

En→Zh Conversation En→Zh Ecommerce Zh→En CommonSense Zh→En Social De→En Conversation De→En Ecommerce

484 530 1200 491 462 501

En→Zh Social En→Zh Culture Zh→En Conversation Zh→En Ecommerce De→En Social

511 778 349 518 515

Table 10: Out-of-domain test sets and sample counts for En↔Zh and De→En translation tasks.

2022). The Culture domain is sourced from the CAMT dataset (Yao et al., 2024), and the CommonSense domain comes from the CommonMT benchmark (He et al., 2020). The data statistics is illustrated in Table 10. B.5

Domain Diversity Design

To ensure comprehensive evaluation and robust generalization of TwT across diverse translation scenarios, we designed the dataset to reflect a broad range of linguistic and contextual complexity, including: high-resource domains (e.g., News); low-resource scenarios (e.g., Koranic texts); terminology-heavy fields (e.g., IT, Law, Medicine, Science, E-commerce); context-sensitive domains

(e.g., Culture, CommonSense); stylistically demanding content (e.g., Literature); noisier or informal genres (e.g., Subtitles, Conversation, Social media). This diversity ensures comprehensive evaluation across domain specificity, reasoning difficulty, and stylistic variation. B.6

Multilingual Test Data

For unseen-language evaluation, we adopt the FLORES+ benchmark (NLLB Team et al., 2024) and construct an unseen language set to minimize evaluation leakage from the languages already involved in our baseline training data (Table 5). Concretely, we first exclude all languages that appear in the baseline training coverage, including Chinese (zh), English (en), German (de), French (fr), Spanish (es), Portuguese (pt), Italian (it), Russian (ru), Korean (ko), Dutch (nl), Czech (cs), Icelandic (is), Ukrainian (uk), Hindi (hi), Japanese (ja), Polish (pl), Swedish (sv), Hungarian (hu), Romanian (ro), Danish (da), Norwegian (no), and Finnish (fi). This step ensures that the En↔X results reflect generalization to genuinely unseen languages rather than memorization of language-specific posttraining signals. Next, to enable consistent computation of COMET and CometKiwi across all unseen directions, we further restrict the remaining FLORES+ languages to those supported by our COMET/CometKiwi scorers. After these two filters, we obtain a final unseen-language set consisting of 59 languages (Table 12). The seen languages—German (de), English (en), and Chinese (zh)—are used across all baseline training sets, and are covered by the follow-

Model

Accuracy (↓)

Style (↓)

Fluency (↓)

Terminology (↓)

Non-translation (↓)

50.55 52.55 50.52 50.13 52.31 49.82 50.68 55.02

30.50 29.55 30.76 30.97 29.70 30.70 29.69 26.37

6.95 6.91 7.11 6.71 6.54 7.13 7.12 6.89

11.90 10.81 11.40 12.10 11.31 12.23 12.37 11.64

0.09 0.18 0.21 0.09 0.13 0.11 0.14 0.08

DeepSeek-V3 Gemini-2.0-Flash GPT-4o DeepSeek-R1 Gemini-2.0-Flash-Thinking OpenAI-o3-mini OpenAI-o1 TwT-Qwen2.5-14B-Instruct

Table 11: MQM-based error analysis across different LLMs and LRMs.

ode

Language

Code

Language

Code

Language

afr asm bos cym epo gle hau hye kan khk lao mal mya pbt sin som swh tha urd ydd

Afrikaans Assamese Bosnian Welsh Esperanto Irish Hausa Armenian Kannada Mongolian Lao Malayalam Burmese Pashto Sinhala Somali Swahili Thai Urdu Yiddish

als bel bul ekk eus glg heb ind kat khm lit mar npi plt slk srp tam tur vie zsm

Albanian Belarusian Bulgarian Estonian Basque Galician Hebrew Indonesian Georgian Khmer Lithuanian Marathi Nepali Malagasy Slovak Serbian Tamil Turkish Vietnamese Malay

amh ben cat ell fil guj hrv jav kaz kir lvs mkd pan san slv sun tel uig xho

Amharic Bengali Catalan Greek Filipino Gujarati Croatian Javanese Kazakh Kyrgyz Latvian Macedonian Punjabi Sanskrit Slovenian Sundanese Telugu Uyghur Xhosa

Table 12: The 59 unseen languages Lunseen used for En↔X evaluation after filtering FLORES+ by (i) posttraining language coverage of evaluated backbones and (ii) COMET/CometKiwi language support.

ing datasets: the German-English multi-domain dataset (Aharoni and Goldberg, 2020), the EnglishChinese UM-Corpus (Tian et al., 2014), and the Chinese-English GuoFeng-Webnovel dataset.

C

Training Data Scale of MT Baselines

To ensure a fair comparison, we report the training data size used by each MT-specialized baseline. While TwT and most reasoning-augmented models, including MT-R1-Zero-7B, CoT-FT-7B, SFT-Parallel, and mExTrans-7B, are trained on approximately 27K examples, several other models leverage significantly larger corpora. For instance, TowerInstruct series is trained on 637K examples, Tower-Plus-9B on 286K, and ALMA-R on 21K. SSR-X-Zero-7B is trained on a notably smaller subset of 13K instances.

D

API Details

The following APIs were used to access the SOTA LLMs and LRMs evaluated in our experiments: • OpenAI: gpt-4o-2024-11-20, o1-2024-12-17, o3-mini-2025-01-31, and gpt-5-2025-08-07 • DeepSeek: deepseek-chat-2024-12-26, and deepseek-reasoner • Gemini: gemini-2.0-flash, and gemini-2.0flash-thinking-exp-01-21

E

Experimental Details for Preliminary Analysis

E.1

Setup for Pure RL

Training Data. We construct a diverse MDMT training data with 20K samples by randomly sampling 2K sentence pairs from 10 domains, spanning three major language pairs: De→En (Aharoni and Goldberg, 2020), En→Zh (Tian et al., 2014), and Zh→En (Wang et al., 2023, 2024a). We select only examples with source sentences longer than 20 words (characters for Chinese) to encourage meaningful reasoning. Evaluation is conducted using the test sets provided by the original datasets. Training Template. To guide the base model toward producing translation-relevant reasoning behavior, we modify the template used in DeepSeekR1-Zero (Guo et al., 2025) to better fit the translation task. The chat template is shown in Figure 5. Reward Design. We follow DeepSeek-R1 (Guo et al., 2025) in using two reward types. Format Reward: We apply regex matching to check whether the model responses adhere to the specified format. The format reward rf is set to 1 if the format is correct, and -1 otherwise. Quality Reward: We experiment with several quality metrics to evaluate the final translation output. Specifically, we consider BLEU6 (Papineni 6

https://github.com/mjpost/sacrebleu

et al., 2002), a metric based on n-gram lexical overlap, and COMET7 (Rei et al., 2020) and CometKiwi8 (Rei et al., 2022), neural metrics that estimate semantic similarity using contextual embeddings through reference-based or reference-free scoring. All metric scores are normalized to the range [0, 1] before being used in the reward function. During training, the quality reward rq is computed based on one or more of these metrics. It is only applied when the output format is correct; otherwise, we assign a fixed penalty of rq = −2. The final reward used during RL training is the sum of the format and quality components: r = rf + rq . RL Optimization. We train the model using GRPO algorithm (Shao et al., 2024), which improves stability over PPO (Schulman et al., 2017) by leveraging group-wise relative advantages rather than explicit value functions. See Appendix A for full formulation. Model Training. We conduct experiments using Qwen2.5-7B-Instruct (Yang et al., 2024) as the base model, and train it for 1 epoch on our collected 20k MDMT data. To investigate the effectiveness of different reward signals, we experiment with five quality reward variants: BLEU, COMET, CometKiwi, BLEU+COMET, and BLEU+CometKiwi.

E.2

Setup for RL with SFT

Training Data. We distill reasoning traces from the DeepSeek-R1 model (Guo et al., 2025) and apply additional quality filtering, resulting in about 7K high-quality Long CoT examples for MDMT. Model Training. We fine-tune Qwen2.5-7BInstruct (Yang et al., 2024) on the constructed Long CoT dataset for 1 epoch, followed by RL training on the 20K multi-domain data for 1 epoch, using a composite reward of BLEU and COMET. General-CoT and Domain-CoT share the same training pipeline and hyperparameters, differing only in the prompt used for SFT data construction: General-CoT uses "Translate the following {src_lang} text into {tgt_lang}.", while Domain-CoT uses "Translate the following {src_lang} text into {tgt_lang} while maintaining the domain style of the source text." 7 8

Unbabel/wmt22-comet-da Unbabel/wmt22-cometkiwi-da

β

BLEU

COMET

KIWI

Quality

Lens.

0 0.001 0.005 0.01 0.02 0.05

30.07 30.58 30.79 30.37 30.24 29.71

82.35 82.52 82.48 82.49 82.37 82.23

81.20 81.21 81.28 81.40 81.29 81.24

64.54 64.77 64.85 64.75 64.63 64.39

130 278 273 282 254 275

Table 13: In-domain results with different KL coefficient.

F

Further Analysis

F.1

Error-Type Analysis under the MQM Framework

To further investigate what types of translation errors reasoning helps address, we conducted a detailed analysis based on the MQM (Multidimensional Quality Metrics) framework. For error classification, we used an external LLM as an annotator, excluding "Source Error" as it pertains to lowquality source data. Given the distinct challenges of MDMT compared to general MT, we extended error types such as "Style Inconsistency", "Crossdomain Confusion", and "Terminology Misuse" to better capture the advantages of reasoning in MDMT. As shown in Table 11, reasoning-enhanced LRMs lead to lower error rates in Style, Fluency, Terminology, and Non-translation Errors. We believe this is due to the structured reasoning process, which contributes to a more human-like translation process and better error reflection, in contrast to the pattern-matching of traditional bilingual finetuning. Notably, after incorporating domain-aware CoT data, TwT reduced style errors by 3.18%, significantly improving the model’s style adaptation. This highlights the effectiveness and applicability of our approach in MDMT. F.2

KL Coefficient Analysis.

Recent study (Yu et al., 2025) suggests that removing KL regularization can enhance Long CoT reasoning. To examine whether this holds in the MDMT setting, we investigate the effect of the KL coefficient β in Table 13. However, our findings diverge from this conclusion: setting β to 0—completely removing the KL constraint—results in significantly shorter and less informative outputs. In contrast, small non-zero values (e.g., β = 0.001 or 0.005) achieve the best overall quality scores while maintaining reasonable response lengths. Larger values (e.g., β = 0.02 or 0.05) overly constrain the

Reward vs Training Step

Format Error Rate vs Training Step

2.1

2.0 TwT-7B TwT-14B

1.9 0

250

500

750

1000

RL Training Steps

1250

Format Error Rate (%)

2.2

Reward

TwT-7B TwT-14B

1.50 1.25 1.00

IT

Prompt

0.75

General prompt Domain-aware prompt

0.50

Law

Medical

BLEU

COMET

KIWI

BLEU

COMET

KIWI

BLEU

COMET

KIWI

36.66 36.55

83.49 83.57

78.68 79.82

39.16 40.57

84.97 85.32

82.35 83.54

40.23 41.14

83.82 84.06

81.81 82.97

0.25 0.00 0

250

500

750

1000

1250

RL Training Steps

Figure 7: Average reward and format error rate over RL training steps.

policy, slightly degrading translation quality and increasing length instability. These results reflect the importance of balancing generation stability and flexibility: moderate KL regularization helps suppress degenerate behavior while preserving adaptive, difficulty-aware reasoning traces. F.3

Table 14: General prompt vs. domain-aware prompt across domains.

Analysis of Training Dynamics.

To better understand how RL shapes TwT’s reasoning and translation behavior, we analyze the training dynamics of TwT-7B and TwT-14B from two complementary perspectives. Analysis of Reward and Format Stability Figure 7 further illustrates the reinforcement learning dynamics of TwT-7B and TwT-14B in terms of average reward and format error rate. Both models show a rapid increase in reward within the first few hundred steps, followed by a stable plateau, indicating that the hybrid reward formulation enables efficient convergence toward high-quality translation reasoning. The larger TwT-14B model consistently achieves higher rewards throughout training, suggesting stronger optimization capacity and better utilization of the reward signal. In contrast, the format error rate decreases sharply during the initial phase and remains near zero thereafter, demonstrating that the format reward effectively reduces structural inconsistencies and output anomalies as training progresses. Translation Quality and Reasoning Depth Figure 8 shows the training dynamics of TwT-7B and TwT-14B during RL. BLEU, COMET, and CometKiwi scores steadily improve, indicating that our hybrid reward effectively enhances translation quality. Interestingly, response length increases in the early and mid stages, reflecting deeper reasoning, but decreases in later steps as the model learns to maintain quality with more concise traces. This indicates convergence to efficient, difficulty-

Table 15: SFT trained on general CoT vs. domain-aware CoT. In-Domain

Method

Out-of-Domain

BLEU

COMET

KIWI

BLEU

COMET

KIWI

SFT w/ general CoT SFT w/ domain-aware CoT

23.51 23.74

80.52 80.74

80.15 80.46

25.12 25.67

83.40 83.57

80.81 81.18

∆

+0.23

+0.22

+0.31

+0.55

+0.17

+0.37

aware reasoning behavior, as shown in Figure 8(a). Figure 8(b) and Figure 8(c) show the performance of TwT-7B and TwT-14B under different difficulty levels. In both models, BLEU, COMET, and CometKiwi scores are highest on easy inputs and lowest on hard ones, and response length increases with difficulty, indicating effective control of reasoning depth. Notably, TwT-14B generates shorter responses than TwT-7B on easy inputs, but longer responses on medium and hard inputs. This suggests that the larger model more effectively adapts its reasoning length to input difficulty, providing concise outputs when possible and allocating more reasoning to harder cases. F.4

Analysis of Domain-Aware Prompting

To better understand the effect of domain-aware prompting, we conduct two complementary experiments that evaluate its impact from both the inference and training perspectives. First, we investigate how prompt formulation affects translation quality when the model remains fixed (DeepSeek-R1). Two prompts are compared: a general prompt ("Translate the following src_lang sentence into tgt_lang.") and a domain-aware prompt ("Translate the following src_lang text into tgt_lang while maintaining the domain style of the source text."). As shown in Table 14, the comparison across three representative domains (IT, Law, Medical) demonstrates that domain-aware prompt explicitly instructs the model to infer and preserve domain-specific stylistic and terminological features. This better activates the reasoning capacity and leads to more accurate domain-aware translation. These findings motivated us to adopt domain-aware prompts throughout TwT’s training data curation phase, allowing domain signals to be explicitly injected dur-

300 250 200

82.5

81.1

30.1

82.0

80.9

28.9

250

500

750

RL Training Steps

1000

26.5

1250

81.5 81.0

0

250

500

750

RL Training Steps

1000

80.5

1250

CometKiwi vs Training Step

81.3

31.3

27.7 0

COMET vs Training Step

83.0

COMET

350

BLEU

Response Length

400

BLEU vs Training Step

32.5

TwT-7B TwT-14B

CometKiwi

Response Length vs Training Step

450

80.7 80.5

0

250

500

750

RL Training Steps

1000

80.3

1250

0

250

500

750

RL Training Steps

1000

1250

(a) Training dynamics for TwT-7B and TwT-14B.

225 175 125

85.2

82.8

30.6

83.4

81.6

28.4

250

500

750

RL Training Steps

1000

24.0

1250

81.6

80.4

79.8 0

250

500

750

RL Training Steps

1000

78.0

1250

CometKiwi vs Training Step

84.0

32.8

26.2 0

COMET vs Training Step

87.0

COMET

275

BLEU

Response Length

325

BLEU vs Training Step

35.0

Easy Medium Hard

CometKiwi

Response Length vs Training Step

375

79.2 0

250

500

750

RL Training Steps

1000

78.0

1250

0

250

500

750

RL Training Steps

1000

1250

(b) Training dynamics for TwT-7B under different difficulty level.

315 220 125

85.2

82.8

30.6

83.4

81.6

28.4

250

500

750

RL Training Steps

1000

1250

24.0

81.6

80.4

79.8 0

250

500

750

RL Training Steps

1000

1250

78.0

CometKiwi vs Training Step

84.0

32.8

26.2 0

COMET vs Training Step

87.0

COMET

410

BLEU

Response Length

505

BLEU vs Training Step

35.0

Easy Medium Hard

CometKiwi

Response Length vs Training Step

600

79.2 0

250

500

750

RL Training Steps

1000

78.0

1250

0

250

500

750

RL Training Steps

1000

1250

(c) Training dynamics for TwT-14B under different difficulty level.

Figure 8: Training dynamics for TwT models.

ing learning. Second, we evaluate the effect of using different prompts to construct Long-CoT data for SFT. Specifically, one dataset is generated using the general prompt, while another uses the domain-aware prompt to elicit domain-specific reasoning traces. As shown in Table 15, domain-aware CoT data lead to consistent improvements across all automatic evaluation metrics in both in-domain and out-of-domain scenarios. These results highlight that incorporating domain context enhances the model’s reasoning robustness and cross-domain generalization. F.5

Computational Cost Analysis

To assess the computational efficiency of our reasoning-based translation paradigm, we compare the average inference time per example across dif-

12

Qwen2.5-7B-Instruct + SFT

57

TwT-7B (ours)

121

TwT-14B (ours) 0

20

40 60 80 100 Time Cost (ms/example)

120

Figure 9: Inference time comparison.

ferent models. As shown in Figure 9, the reasoningaugmented TwT models incur additional computational overhead compared with the non-reasoning baseline (Qwen2.5-7B-Instruct + SFT). Specifically, TwT-7B and TwT-14B require 57 ms and 121 ms per example, respectively, compared to 12 ms for the baseline. Despite the higher inference cost, the time overhead remains acceptable given the

significant improvements in translation quality, reasoning accuracy, and style adaptation demonstrated in previous sections. F.6

Language Consistency Discussion

To verify that TwT’s reasoning traces are linguistically coherent and aligned with the target translation language, we conducted a systematic analysis of language consistency throughout the data curation and training stages. We employed a language identification tool (fastText) to detect potential cross-lingual inconsistencies in reasoning traces, such as mixing English reasoning with Chinese or German summaries. In the initial R1-distilled CoT data, 9.26% (684/7382) of the reasoning traces contained incorrect languages. After GPT-4o rewriting, the inconsistency dropped sharply to 3.78% (279/7382). After SFT, the inconsistency rate on the test set further decreased to 0.04% (7/16094), and it reached 0% after the RL stage. F.7

Reliability of Difficulty Estimation

To mitigate potential bias from relying solely on GPT-4o, we validated its difficulty estimation against two other strong models (DeepSeek-V3.2, Gemini-2.5-Flash) and human experts on 100 randomly sampled instances. Results demonstrate robust consistency across three dimensions: (1) Internal Stability: All models exhibited high selfagreement across 5 independent trials (GPT-4o: 0.92, DeepSeek: 0.93, Gemini: 0.91). (2) CrossModel Agreement: The three models reached a 90% consensus rate after majority voting. (3) Human Alignment: Crucially, GPT-4o achieved the highest correlation with professional translators (0.86), surpassing DeepSeek (0.77) and Gemini (0.79). These findings confirm that GPT-4o provides a reliable and stable proxy for human judgment in our difficulty-aware framework. F.8

Structured Analysis of Failure Cases

To provide a more systematic analysis, we further examine TwT’s bad cases across all domains. Specifically, we select the 10 lowest-quality examples from each of the 15 domains (150 cases in total) and analyze their CoT trajectories. We identify two recurring error patterns. Reasoning–prediction misalignment. In some cases, the model makes the correct translation decision in its reasoning, but the final output fails to realize it. That is, the reasoning identifies an appro-

Human Evaluation TwT-Qwen2.5-14B-Instruct vs. DeepSeek-R1 (671B) TwT-Qwen2.5-14B-Instruct vs. DeepSeek-V3 (671B) TwT-Qwen2.5-14B-Instruct vs. TowerInstruct-13B-v0.1

TwT Win

TwT Lose

Tie

0.41 0.275 0.54

0.47 0.405 0.30

0.12 0.32 0.16

Table 16: Summary of human evaluation results for TwT compared with other models.

priate lexical choice or phrasing, while the prediction deviates from it. This suggests a gap between reasoning and realization. A possible remedy is to introduce an additional reward penalty during GRPO to explicitly discourage divergence between the decision expressed in the <think> block and the output in the <answer> block. Domain–terminology misalignment. Another common failure occurs when the model correctly identifies the domain but does not switch to the appropriate terminology system. In such cases, the reasoning remains at a declarative level without consistently applying domain-specific lexical choices, syntactic preferences, fixed translations, or formatting conventions. One possible remedy is to incorporate external terminology lexicons or style-template libraries to guide domain-specific realization.

G

Human Evaluation

G.1

Human Evaluation Analysis

We conducted a human evaluation to complement our automatic metrics. Specifically, we randomly sampled 100 examples from the Zh↔En test set, selecting 20 sentences from each of five domains. For each instance, the source sentence and two system outputs—one from TwT-14B and one from a comparison model—were independently evaluated by three professional translators, who selected the better translation or marked a tie when the quality difference was negligible. Table 16 summarizes the human evaluation results across three baselines: DeepSeek-R1, DeepSeekV3, and TowerInstruct-13B-v0.1. These results show that while TwT-14B slightly underperforms DeepSeek-R1 and DeepSeek-V3—expected given its much smaller size—it still achieves a strong degree of parity. Importantly, TwT-14B outperforms TowerInstruct-13B-v0.1 in more than half of the evaluated examples (54% win rate), supporting the effectiveness of its reasoning-driven design in translation.

G.2

Comparison with Human Translation Reasoning

To further evaluate the alignment between TwT’s reasoning process and that of professional human translators, we conducted a case study using a complex narrative sentence from a fictional novel, as shown in Figure 17. The source sentence features multiple subordinate and concessive clauses, complex logical progression, and dense world-specific terminology—a typical example in the fictional novel domain. Such sentences require precise control of syntax, consistent terminology management, and sensitivity to logical flow and pragmatic tone. Similarities. TwT demonstrates several key reasoning behaviors consistent with professional translators. First, it correctly identified the sentence structure and logical relations, decomposing the original into two conceptual layers: (1) background explanation of the connection between the 神属大 世界 and the 战星联邦, and (2) a reasoning-based denial of the assumed link between the 组织 and 神 属大世界. This decomposition mirrors the syntactic and logical analysis stage in human translation reasoning, reflecting TwT’s strong ability to detect hierarchical structure and causal relationships. Second, TwT shows strong terminological consistency, accurately translating key domain-specific entities such as "Immortal Holy Water" (不老圣水), "Divine Realm Major World" (神属大世界), and "War Star Federation" (战星联邦). While minor stylistic improvements remain possible, the model’s use of consistent and semantically clear terms aligns with the terminology standardization step in professional translation workflows. TwT also handled subtle pragmatic expressions such as "有交情" and "换取" with contextually appropriate equivalents ("have connections with", "trade for"), indicating an emerging awareness of functional-pragmatic equivalence—a key element in expressing communicative intent and contextual tone. Moreover, before translating, TwT demonstrated an initial assessment of text difficulty and style, recognizing that the sentence belongs to a fictional novel with specialized terms. This awareness parallels a human translator’s pre-translation difficulty assessment and background analysis, where genre and register are evaluated to inform strategy. TwT chose to preserve the source-style transliteration rather than domesticate the names, maintaining consistency with the fantasy universe’s internal logic—an appropriate decision for this genre. At the structural level,

TwT effectively reconstructed the concessive–causal logic of "虽然. . . 但. . . 所以. . . " into the English pattern "Although...some...therefore...", faithfully capturing the original logical progression. This demonstrates the model’s ability to rebuild syntactic and logical relations during target-language reorganization, consistent with human translators’ reasoning in structural decomposition and coherence reconstruction. In lexical judgment, TwT made contextually informed choices, such as rendering "不老圣水" as "Immortal Holy Water" rather than the freer "elixir of youth" or "anti-aging water", thereby preserving the source’s mythological tone. Similarly, translating "牵强" as "far-fetched" appropriately conveys the intended skepticism while maintaining stylistic naturalness. These decisions illustrate the model’s ability to perform semantic disambiguation and lexical selection comparable to human translators’ third-stage reasoning process. Finally, TwT’s output features a natural syntactic flow and coherent discourse structure. For instance, "some powerful individuals...had connections with..." reconstructs the information hierarchy more fluently than a literal rendering would. This indicates an emerging sense of register and stylistic adaptation, partially fulfilling the requirements of the expression and style selection stage. Differences. First, TwT’s translation repeatedly uses the long-form expression such as "Divine Realm Major World", resulting in verbosity. Professional translators would typically balance terminological consistency with referential economy, adopting simplified references such as "the Divine Realm" or "that realm" to improve fluency and readability. For instance, a revised version could read: "Although ‘Immortal Holy Water’ is a specialty of the Divine Realm, some powerful individuals in the War Star Federation have connections with that realm and can obtain it through trade." This alternative maintains precision while achieving smoother rhythm and reduced redundancy. Second, TwT’s stylistic control remains limited. Although the translation is grammatically accurate, it lacks the conversational tone and personality expected in dialogue. Since the original sentence is dialogue from Hans expressing skepticism, a professional translator would employ a more natural, idiomatic style to reflect the speaker’s voice and pragmatic intent, e.g., "Come on, you can’t just assume they’re connected to the Divine Realm over a bit of Holy Water—that’s a stretch!" Such adaptation better re-

produces conversational tone and aligns with the narrative style of fantasy dialogue. Third, TwT’s cultural adaptation is surface-level. While "farfetched" effectively captures the core meaning of "牵强," a human translator might enrich cultural accessibility by appending a light explanatory note, such as "Immortal Holy Water, a sacred elixir rumored to halt aging in the Divine Realm." This hybrid "foreignization plus domestication" strategy balances cultural retention with reader comprehension, commonly seen in professional literary translation.

H

Prompt

The template for pure RL (Section 3.1) is shown in Figure 5. The prompt for generating difficultyadaptive Long CoT data is provided in Figure 10. The prompt employed for the quantitative evaluation of reasoning efficiency is presented in Figure 12. The prompt for evaluating translation difficulty is shown in Figure 11 and the detailed examples for translation difficulty evaluation are provided in Table 17.

I

Detailed Results by Metric

For completeness, we provide the full breakdown of BLEU, COMET, and CometKiwi for the results reported in Table 3, Table 4, and Table 5. The corresponding detailed results are shown in Table 18, Table 19, and Table 20, respectively.

You are an experienced translation expert. Your task is to optimize a given initial translation thought process by simulating the thinking process of a professional human translator. Human translators typically first assess the difficulty of the translation task. this difficulty level, they allocate appropriate time and cognitive effort:

Based on

• For simple translation tasks, they provide an accurate translation quickly with concise thought process. • For moderate translation tasks, they reason carefully through any ambiguities to produce an accurate translation, without excessive expansion. • For difficult translation tasks, they conduct deep and comprehensive thinking, exploring different translation strategies, comparing and verifying options, and refining every detail to produce the best translation. Based on this professional approach, please optimize the given translation thought process by: • Integrating the assessment of translation difficulty naturally into the thought process. • Adjusting the depth and style of reasoning based on the difficulty level. • Preserving the original language style and tone of the initial translation thought process. Finally, output the optimized thought process and the final translation in JSON format with the keys "thought" and "translation", as shown below: { "thought": "Optimized translation thought process based on difficulty", "translation": "Final translation based on the corrected thought process" } Source text: Target language: Initial translation thought process:

Figure 10: Full prompt used for generating difficulty-adaptive Long CoT data with GPT-4o.

Source Sentence

Reference Translation

Difficulty

Reason

“我只是想要听听您的意见。”她 用一副恭敬的口吻说道。

“I just want to hear your opinion,” she said in a respectful tone.

Easy

A simple sentence with everyday vocabulary and direct grammar. No specialized knowledge required.

At the same time the waves are fanning out, they are also separating by wavelength, a process known as dispersion.

当这些波向外扩散时，它们也 在按波长分开，这一过程叫频 散。

Medium

Contains a compound sentence and one scientific term (*dispersion*), requiring moderate structural and lexical control.

Die folgende Tabelle zeigt die Häufigkeit und Schwere unerwünschter Wirkungen, die in >5% von 168 Patienten mit Mesotheliom berichtet wurden...

The table below provides the frequency and severity of undesirable effects reported in >5% of 168 mesothelioma patients...

Hard

Long and syntactically complex, includes multiple subordinate clauses, verb-final word order, and domain-specific medical terminology. Requires structural parsing and domain expertise.

Table 17: Detailed examples for translation difficulty evaluation.

Your task is to assess the difficulty of translating a given {src_lang} sentence into {tgt_lang}. Please evaluate the difficulty based on the following criteria: 1. Sentence complexity: Determine if the sentence is a simple sentence, a compound sentence, or includes subordinate clauses and other complex structures. 2. Vocabulary difficulty: Assess whether the sentence contains commonly used words or specialized terms or slang. 3. Grammar differences: Analyze if the sentence’s grammatical structure is similar to or differs significantly from {tgt_lang}. 4. Contextual understanding: Consider whether understanding specific cultural contexts or background knowledge is necessary for accurate translation. The difficulty level should be rated as "easy", "medium", or "hard". Additionally, provide a brief, simple reason for the assigned difficulty level. Output the result in JSON format with the keys "level" and "reason", as follows. { "level": "easy/medium/hard", "reason": "simple explanation of the difficulty level." } Here is the {src_lang} sentence: {src_text}

Figure 11: Full prompt used for evaluating translation difficulty with GPT-4o.

You are an expert model specialized in “Chain-of-Thought (CoT) Quality Evaluation for Multi-Domain Translation Tasks.” Your task is to compare the reasoning traces of Model A and Model B: Identify redundancy types in Model A’s reasoning and determine whether Model B has successfully eliminated these redundancies. Please strictly adhere to the following Redundancy Type Definitions for your analysis. If you identify a new redundancy type in Model A, you may include it in the output with a supplemental explanation. [Multi-Domain Translation: Redundancy Definitions] • Over-segmentation: Excessively decomposing simple sentences or obvious content into word-by-word or phrase-by-phrase fragments, causing unnecessary step inflation. • Unnecessary linguistic explanation: Over-explaining common sense or obvious linguistic points (grammar, part-of-speech, etymology, etc.) that do not influence the translation outcome. • Semantic repetition: Repeating the explanation of the same meaning or using different expressions to illustrate the same semantic point without adding value to the translation. • Irrelevant information: Introducing background knowledge, domain trivia, speculative content, or information unrelated to translation decisions (e.g., unnecessary context assumptions). • Redundant alternative translations: Providing multiple translation candidates with highly similar meanings and low contribution, or performing iterative optimization on trivial differences. • Low-density long descriptions: Using verbose sentences to express simple content or using a large volume of text to derive obvious reasoning steps. [Task Requirements] • Identify redundant snippets in Model A’s CoT and label them with the corresponding redundancy type. • Explain why the snippet belongs to this category. • Determine if Model B’s CoT eliminated this specific redundancy (Yes/No). • Output strictly as a structured JSON array in the following format: [ { "type": "Name of the redundancy type", "before_snippet": "Text snippet from Model A", "reason": "Reason for the judgment", "after_resolved": "Yes/No" }, ... ] Now, please analyze the following two reasoning chains: Model A CoT: {model_a_cot} Model B CoT: {model_b_cot}

Figure 12: The prompt used for the quantitative evaluation of reasoning efficiency.

Source Text: Oh, ich kann nicht glauben, dass sie weggelaufen ist.

Long CoT data distilled from DeepSeek-R1 (562 tokens): <think>Okay, let's tackle this translation task. The user wants the German sentence "Oh, ich kann nicht glauben, dass sie weggelaufen ist." translated into English. First, I need to understand the sentence. The main components are "Oh, I can't believe that she ran away." The speaker is expressing surprise or disbelief that someone (female) has run away. Next, the user mentioned identifying the domain and ensuring the translation reflects the domain's style. Hmm, domains could be things like literature, casual conversation, formal reports, etc. The original sentence uses "Oh" and "kann nicht glauben," which are pretty colloquial. It sounds like conversational or maybe even a dialogue from a novel or a TV show. The emotional tone here is informal and expressive. So the domain is likely casual or conversational. Maybe even a dramatic context, like a soap opera or a novel. Now, translating it to match that style. The direct translation would be "Oh, I can't believe she ran away." But depending on the domain, maybe we can make it a bit more natural. For example, in casual speech, people might contract "cannot" to "can't," which is already done. "Weggelaufen" is "ran away," which is correct. Wait, the original uses "sie," which could be "she" or "they" depending on context. But since the user provided the German sentence, and the verb is "weggelaufen ist" which is singular (since "ist" is singular), so "she" makes sense here. Unless the context is formal, but the sentence seems informal. So "she" is correct. Also, the structure "dass sie weggelaufen ist" is a subordinate clause. In English, "that she ran away." But in conversational English, sometimes people drop "that," so "Oh, I can't believe she ran away." That sounds more natural and fits the casual domain better. So the domain here is probably everyday conversation or a dramatic dialogue. The style should be informal, using contractions and natural phrasing. So the translation should reflect that. No need for formal language. I think that's it. Let me double-check. The original is expressing surprise, so the translation needs to keep that emotional tone. "Oh" at the beginning sets that. "Can't believe" is better than "cannot believe" for informality. Yes. So the final translation would be "Oh, I can't believe she ran away." And the domain is casual conversation or possibly a narrative in fiction. The style is informal and expressive.</think> <answer>Oh, I can’t believe she ran away.</answer>

Our curated difficulty-adaptive Long CoT data (167 tokens): <think>First, I assess the difficulty of the translation: this is a simple, conversational German sentence expressing surprise or disbelief about someone‘s action. There are no idioms or complex structures, so the translation is straightforward. The sentence is “Oh, ich kann nicht glauben, dass sie weggelaufen ist.”—literally, “Oh, I can’t believe that she ran away.” To preserve the informal and expressive tone, it‘s best to use the contraction “can’t” and drop the explicit ‘that’ for natural conversational English. ‘Oh’ at the start effectively conveys the sense of surprise. No ambiguities or special cultural references require deeper reasoning. Thus, the optimal translation is clear and concise.</think> <answer>Oh, I can‘t believe she ran away.</answer>

Source Text: 现在他要找的就是外门弟子所修炼的地方。

Long CoT data distilled from DeepSeek-R1 (476 tokens): <think>好的，我现在要处理这个翻译任务。用户给的中文句子是：“现在他要找的就是外门弟子所修炼的地方。”首先，我需要确定这句话的领域。看起来像是武侠小说 或者修仙类的小说，因为提到了“外门弟子”和“修炼的地方”，这些词汇常见于这类文学作品中。 接下来，我需要确保翻译符合目标领域的风格。武侠或修仙小说通常使用较为正式且带有古风的英语表达，同时保留一些特定的术语。比如“外门弟子”通常翻译为 “outer sect disciples”，而“修炼的地方”可以是“training grounds”或者“practice area”。这里可能需要根据上下文选择更合适的词汇，但“training grounds”听起来比 较自然。 然后，我需要分析句子的结构。原句的结构是“现在他要找的就是……的地方”，强调当前的目标。翻译时要保持这种强调，可能需要用“What he needed to find now was...”这样的结构，这样既符合英文表达习惯，又传达了原句的重点。 另外，时态方面需要注意。原句中的“现在”可能指的是当前的情况，用过去时的话可能更适合叙述故事，所以用“needed”比较合适，因为很多小说是用过去时叙述的。 不过如果是现在进行的情节，可能用现在时，但根据上下文，假设是过去时更合适。 还要确保术语的一致性。比如“外门弟子”在整个作品中可能有固定的翻译，所以保持一致性很重要。如果用户之前有特定的译法，可能需要调整，但这里没有其他信息， 所以采用通用的译法。 最后，检查整个句子的流畅性和准确性。确保没有遗漏信息，并且符合目标领域的风格。比如是否需要用更古风的词汇，或者是否需要调整语序来更符合英语习惯。比如 原句中的“所修炼的”可以用定语从句处理，放在“地方”后面，这样更自然。 总结一下，翻译后的句子应该是：“What he needed to find now was the training grounds of the outer sect disciples.” 这样既准确传达了原意，又符合武侠或修仙小说的风格。 </think> <answer>What he needed to find now was the training grounds of the outer sect disciples.</answer>

Our curated difficulty-adaptive Long CoT data (325 tokens): <think>在接到这个翻译任务时，我首先要对原句“现在他要找的就是外门弟子所修炼的地方。”进行难度评估。从内容来看，这是一句来自武侠或修仙小说的叙述句，涉 及专有名词“外门弟子”和“修炼的地方”，但整体语法结构和表达并不复杂。因此，我将其归为一般难度的文学文本，需要注意术语准确和风格匹配。 具体处理步骤如下： 1. 理解背景：确认“外门弟子”为宗门中层级较低的弟子，常见翻译为“outer sect disciples”；“修炼的地方”可译为“the place where they cultivate”或更自然的 “training grounds”。 2. 语法与时态：句首“现在”表明当前动作，但考虑到小说常用过去时，主句采用了“needed to find”，更符合英语叙述习惯。 3. 保持强调：原文强调“现在他要找的就是……”，英文中可通过“what he needed to find now was...”或倒装强调，也可简洁表达为“Now, he needed to find the...”。 4. 风格：保持叙述的剧情感和正式度。 结合以上分析，最终优化的translation为“Now, he needed to find the place where the outer sect disciples cultivated.” 这样既准确传递了原文语境和术语，也兼顾了英文的流 畅和符合武侠小说风格。</think> <answer>Now, he needed to find the place where the outer sect disciples cultivated.</answer>

Figure 13: Examples of our curated difficulty-adaptive Long CoT data (Easy and Medium).

Source Text: During a meeting with chargéd 'affaires ad interim Li Zhimin of the Chinese Embassy in Vietnam on 15 June 1956, Vice Foreign Minister of the Democratic Republic of Vietnam Ung Van Khiem solemnly stated that, "according to Vietnamese data, the Xisha Islands and Nansha Islands are historically part of Chinese territory." Le Loc, Acting Director of the Asian Department of the Vietnamese Foreign Ministry, who was present, specifically cited Vietnamese data and pointed out that, "judging from history, these islands were already part of China at the time of the Song Dynasty."

Long CoT data distilled from DeepSeek-R1 (648 tokens): <think>好的，我需要把这段英文翻译成中文。首先，先通读一遍原文，理解内容。 这段文字是关于1956年越南民主共和国副外长与中国驻越南使馆临时代办会面的记录，涉及西沙和南沙群岛的主权归属问题。显然，这属于外交或历史领域的文本，可能 涉及法律或国际关系方面的术语。 接下来，确定领域。这里涉及的是历史事件和外交声明，属于国际关系或历史文献的领域。因此，翻译时需要保持正式、严肃的语气，使用准确的官方术语，比如“临时 代办”、“副外长”、“郑重声明”等。 然后处理具体的翻译点： 1. "chargéd'affaires ad interim" 正确翻译是“临时代办”，注意头衔的准确。 2. "solemnly stated" 要译为“郑重声明”，符合外交场合的正式用语。 3. "Vietnamese data" 翻译为“越南方面的材料”比较合适，避免直译成“数据”。 4. "Acting Director" 是“代理司长”。 5. "cited Vietnamese data" 这里可能需要译为“列举了越南方面的材料”，保持上下文一致。 6. "Song Dynasty" 是“宋朝”，注意朝代名称的正确翻译。 另外，要注意时间、职务和人名的正确翻译。比如“Ung Van Khiem”应译为“雍文谦”，“Le Loc”译为“黎禄”。需要确认这些越南人名的标准中文译名，可能需要查阅 相关资料确保准确性。 在句式结构上，英文中的时间状语从句和直接引语的结构要转换为中文习惯的表达方式，比如将“During a meeting on 15 June 1956”译为“1956年6月15日，……会晤时”， 并将直接引语用引号正确标出。 最后检查整个翻译是否流畅，是否符合外交文献的正式风格，术语是否一致，人名、地名、机构名是否准确无误。比如“亚洲司代理司长黎禄”是否正确，是否有遗漏的 职务或头衔。 可能还需要注意一些细节，比如“西沙群岛”和“南沙群岛”的正确译名，以及“历史资料”和“根据历史来看”等表达是否准确传达原意。 总结来说，整个过程需要准确理解原文内容，确定合适的领域和风格，选择合适的术语和句式，确保翻译的准确性和专业性。</think> <answer>1956年6月15日，越南民主共和国外交部副部长雍文谦在与中国驻越南大使馆临时代办李志民会晤时郑重表示，“根据越南方面的材料，西沙群岛和南沙群岛从 历史上看属于中国领土”。在场的越南外交部亚洲司代理司长黎禄特别列举了越南方面的材料指出，“从历史上看，这些岛屿早在宋朝时期就已属中国”。</answer>

Our curated difficulty-adaptive Long CoT data (550 tokens): <think>首先对翻译难度进行评估：本段原文内容涉及1950年代中越之间的外交历史事件，主体为正式外交辞令和历史归属表述，内容较为严肃，需要在译文中准确还原称 谓、人名、地名，确保外交和官方用语的严谨性。人名、职务译名需要查证权威译法，同时还要处理引号、直接引语及时间状语的顺畅转化。因此，属于较高难度的历史 外交领域翻译任务，需要分步细致推敲。 首先，通读原文，精准捕捉每个时间点、人物、头衔及关键论述。 1. “chargé d'affaires ad interim”标准译法为“临时代办”；“vice foreign minister”译为“副外长”，而越南民主共和国副外长通用译名为“雍文谦”；中国驻越南使馆临时 代办“李志民”。 2. 时间状语“During a meeting ... on 15 June 1956”应译为“1956年6月15日，在中国驻越南使馆临时代办李志民的会见中”。 3. 语气要郑重，外交辞令使用“郑重表示”。 4. “Vietnamese data”一般译为“越南方面资料”； 5. “acting director”常见译法为“代司长”，“Le Loc”应对照权威译名为“黎禄”。 6. “judging from history”——译为“从历史上看”或“据史可证”，结合上下文选用更习惯的表达。 7. 人名和地名要核查官方标准译名。 8. 引号方面，层层嵌套时需注意区分人物直述与转述，建议采用中文标准引号和冒号。译文顺序可适当调整以更贴近中文表达习惯。 完成初稿后，逐句核对语气严谨、表达完整，查漏补缺，确保术语统一和风格正式。再次复核关键历史用词、机构名称准确无误，并确保译文逻辑通顺。</think> <answer>1956年6月15日，越南民主共和国外交部副外长雍文谦在会见中国驻越南使馆临时代办李志民时郑重表示：“根据越南方面的资料，西沙群岛和南沙群岛自古以 来就是中国的领土。”当时在场的越南外交部亚洲司代司长黎禄还专门引用了越南方面的资料指出：“从历史上看，这些群岛在宋代时已经属于中国。”</answer>

Figure 14: Examples of our curated difficulty-adaptive Long CoT data (Hard).

Prompt: Translate the following sentence from Chinese into English.

“我们回去！”

<think>First, I assess the difficulty of this translation. The source sentence “我们回去！” is short and clear, with no cultural nuances or domain-specific terms, making it a simple task. Next, I analyze each component: “我们” translates directly as “we”; “回去” means “go back”, and in English, “Let's go back!” is accurate and natural. Considering possible contexts (such as action movies, novels, or casual conversation), the phrase should maintain its brevity and emphasis. Since there are no special cultural references or ambiguities, a concise and faithful translation is appropriate. Overall, the optimal approach is to translate word-for-word accurately and preserve the original tone.</think> <answer>"Let's go back!"</answer>

153 tokens

Prompt: Translate the following sentence from German into English.

Neuen Eintrag hinzufügen

<think>First, I assess the difficulty of translating "Neuen Eintrag hinzufügen" from German to English. The sentence consists of a straightforward phrase with no cultural nuances or domain-specific jargon, making it a simple task. Next, I analyze each component: "Neuen" translates directly as "new"; "Eintrag" means "entry"; and "hinzufügen" is accurately rendered as "add". In technical contexts, such as software interfaces, precision and brevity are valued, but there is no need for elaborate strategies. Reviewing possible alternatives like "Add new entry", "Add new item", or "Add entry", I find that "Add new entry" maintains accuracy and fidelity to the original while fitting typical UI conventions in English. Overall, this translation requires minimal cognitive effort beyond accurate word choice and attention to register.</think> <answer>Add new entry</answer>

Figure 15: Output examples of our TwT-7B.

182 tokens

Prompt: Translate the following sentence from Chinese into English. 身穿皮甲的中年男子站在甲板旁，向下俯瞰着这一望无际的草原与美景。

<think>First, I assess the difficulty of this translation. The sentence involves literary description and requires attention to accuracy in character details and imagery, but there is no deep cultural nuance or technical challenge, so I consider it a moderate task. Next, I analyze each component: '身穿皮甲的中年男子' translates as 'a middle-aged man wearing leather armor', accurately conveying both the attire and age. '站在甲板旁' becomes 'stood on the deck', which maintains fidelity and fits typical novelistic style. '向下俯瞰着这一望无际的草原与美景' means looking down at the endless grassland and beautiful scenery below; 'looked down at the vast grassland and the beautiful scenery below' preserves the original's emphasis and flow. To optimize, I ensure that 'vast grassland' is appropriately rendered as 'the endless prairie' for naturalness and accuracy. Overall, the sentence should be concise yet vivid, suitable for an English-language novel. After reviewing possible alternatives, such as 'a middle-aged man in leather armor stood on the deck, looking down at the endless grassland and the beautiful scenery below', I confirm that this translation maintains accuracy, fidelity, and appropriate tone.</think> <answer>A middle-aged man in leather armor stood on the deck, looking down at the endless grassland and the beautiful scenery below.</answer>

278 tokens

Prompt: Translate the following sentence from German into English. ii) von Rückständen dieser Stoffe in lebenden Tieren, deren festen und fluessigen Ausscheidungen sowie im Tiergewebe und in tierischen Erzeugnissen wie Fleisch,

Milch, Eiern und Honig beschreiben;

<think>First, I assess the difficulty of this translation. The sentence involves precise legal terminology and requires accuracy in mapping each phrase to its appropriate English equivalent, but there is no unusual cultural nuance or ambiguity, making it a moderate task. Next, I analyze each component: 'von Rückständen dieser Stoffe' translates as 'of residues of these substances', which is accurate and faithful to the source. 'In lebenden Tieren' becomes 'in live animals', a direct and correct choice. 'Ihre festen und fluessigen Ausscheidungen' refers to solid and liquid excreta, accurately rendered as 'their solid and liquid excreta'. 'Sowie im Tiergewebe' means 'and in animal

tissues', with 'animal tissue' being an appropriate and concise term. Finally, 'und in tierischen Erzeugnissen wie Fleisch, Milch, Eiern und Honig beschreiben' should become 'describe in animal products such as meat, milk, eggs and honey'; here, 'products' is suitable and maintains precision, though 'derived products' could also be considered if broader context demands it, but given the specificity of the list, 'animal products' is accurate and clear. Overall, the translation should preserve accuracy and flow typical for formal regulatory language.</think> <answer>(ii) the residues of such substances in live animals, their solid and liquid excreta, in animal tissues and in animal products such as meat, milk, eggs and honey;</answer>

Figure 16: Output examples of our TwT-7B.

310 tokens

Prompt: Translate the following sentence from Chinese into English. “虽然‘不老圣水’是神属大世界的特产，但在战星联邦的一些强者与神属大世界的强者 也有交情，也会换取一些‘不老圣水’，所以单是这点‘不老圣水’就认定组织与神属大 世界有关，就有些牵强了！”汉斯摇头说道。

<think>First, I assess the difficulty of this translation. The source sentence involves domain-specific terms from fictional novels, such as '不老圣水', '神属大世界', and '战星联邦', requiring moderate to high accuracy in terminology and logical reasoning. Therefore, it is a hard task suitable for careful attention to accuracy and flow. Next, I analyze each component: '不老圣水' should be rendered as 'Eternal Youth Holy Water'; '神属大世 界' refers specifically to a 'Divine Realm Major World', which can be translated as 'Divine Realm World' or 'Divine Realm Realm', with both accurate and concise options available; '战星联邦' is accurately translated as 'War Star Federation'. For the phrase '在战星联邦的一些强者与神属大世界的强者也有交情， 也会换取一些‘不老圣水’', I need to ensure that the logic flows smoothly in English, suggesting a connection between the powerful figures from the War Star Federation and those from the Divine Realm World, who might exchange some 'Eternal Youth Holy Water'. This could be expressed as 'Some powerful individuals in the War Star Federation have connections with the powerful beings in the Divine Realm World, and they would also trade for some "Eternal Youth Holy Water"', maintaining accuracy and fidelity. Finally, '单是这点“不老圣水”就认定组织与神属大世界有关，就有些牵强了！' means 'Just because there is a little bit of "Eternal Youth Holy Water", it's too strong to conclude that the organization is related to the Divine Realm World!', emphasizing skepticism. A natural rendering would be: 'However, since "Immortal Holy Water" is a specialty of the Divine Realm World, but some powerful individuals in the War Star Federation also have connections with the powerful beings in the Divine Realm World, and they would also trade for some 'Immortal Holy Water', so it's a bit too far-fetched to conclude that the organization has anything to do with the Divine Realm World just based on this small amount of 'Immortal Holy Water'! Overall, the translation requires precision and logical coherence, fitting the style of a novel.</think> <answer>"Although the ‘Immortal Holy Water’ was a specialty of the Divine Realm Major World, some powerful individuals in the War Star Federation also had connections with the powerful beings in the Divine Realm Major World, and they would also trade for some ‘Immortal Holy Water’. Therefore, it was a

bit too far-fetched to conclude that the organization was related to the Divine Realm Major World just because of this small amount of ‘Immortal Holy Water’!” Hans shook his head and said.</answer>

556 tokens

Figure 17: Output examples of our TwT-7B.

Laws

Method BLEU

COMET

News KIWI

Tokens

BLEU

DeepSeek-V3 Gemini-2.0-Flash GPT-4o

59.58 56.43 48.50

89.15 88.90 88.28

84.42 84.39 84.52

-

35.86 36.39 34.58

DeepSeek-R1 Gemini-2.0-Flash-Thinking OpenAI-o3-mini OpenAI-o1 GPT-5 QwQ-32B

59.34 55.75 42.73 48.11 54.06 43.77

89.15 88.61 87.67 88.32 89.34 87.35

84.68 84.27 84.44 85.13 85.43 84.39

577 702 428 478 784 667

33.35 33.90 32.90 33.93 35.05 33.24

SFT-Parallel-7B ALMA-7B-R ALMA-13B-R TowerInstruct-7B-v0.2 TowerInstruct-13B-v0.1 CoT-FT-7B MT-R1-Zero-7B SSR-X-Zero-7B mExTrans-7B

56.64 36.34 40.33 50.53 52.28 57.05 35.49 38.88 38.77

88.82 85.45 86.89 88.25 88.57 88.76 86.78 86.44 87.11

84.28 81.86 83.10 82.96 83.09 84.35 84.51 83.44 84.44

51 72 56 597

28.89 24.53 26.87 30.72 32.24 30.17 31.69 28.14 25.05

TwT-Qwen2.5-7B-Instruct TwT-Qwen2.5-14B-Instruct

52.56 56.03

88.76 89.07

84.73 84.65

310 320

33.63 34.57

Literary

Method

COMET

Science KIWI

Tokens

Large Language Models 86.94 85.44 86.37 85.14 86.22 85.02 Large Reasoning Models 86.67 85.40 498 86.02 84.79 1149 86.29 85.35 443 86.61 85.47 408 86.65 85.52 740 85.77 84.63 584 MT-Specialized Models 85.43 83.91 83.95 81.64 84.65 82.43 84.61 82.45 85.36 83.09 85.23 83.31 42 86.01 84.54 64 85.87 83.98 52 86.07 85.34 553 Our Models 86.75 84.88 311 86.65 84.63 285

IT

Subtitles

BLEU

COMET

KIWI

Tokens

BLEU

COMET

KIWI

Tokens

33.54 33.47 32.09

88.28 87.69 87.34

85.46 85.24 84.84

-

23.66 24.64 24.91

82.44 81.91 81.91

82.69 82.33 81.93

-

31.82 31.98 31.38 32.32 32.21 32.05

88.08 87.47 87.42 87.92 88.06 87.11

85.47 85.06 85.24 85.56 85.47 84.42

478 1092 385 367 606 563

20.50 23.83 22.95 22.96 23.23 22.58

82.10 80.90 81.90 82.31 82.32 81.54

82.84 81.22 82.40 82.60 82.68 81.29

514 708 355 340 519 584

29.28 22.16 24.37 27.76 29.43 28.72 29.45 28.54 25.74

86.19 84.33 85.72 85.60 86.00 86.02 86.87 87.11 87.34

83.68 81.83 82.59 82.99 83.17 83.48 84.61 84.65 85.31

39 61 49 546

26.93 18.56 19.54 22.45 23.74 28.17 22.48 23.18 14.31

81.41 80.18 81.11 80.75 81.19 80.92 81.73 80.50 81.40

80.26 79.40 80.00 80.01 80.74 79.20 81.70 81.99 82.85

29 55 39 476

32.41 32.75

87.53 87.54

84.77 84.66

294 272

24.74 24.44

82.63 82.73

81.73 81.74

247 241

Koran

Medical

Average

BLEU COMET KIWI Tokens BLEU COMET KIWI Tokens BLEU COMET KIWI Tokens BLEU COMET KIWI Tokens BLEU COMET KIWI Tokens DeepSeek-V3 Gemini-2.0-Flash GPT-4o

16.14 18.37 17.75

77.28 77.20 77.48

77.07 76.67 77.18

-

38.16 37.93 37.23

83.90 83.20 83.52

78.86 78.49 78.37

DeepSeek-R1 Gemini-2.0-Flash-Thinking OpenAI-o3-mini OpenAI-o1 GPT-5 QwQ-32B

11.25 18.01 17.48 16.54 15.01 12.99

75.44 77.06 76.77 77.50 76.72 75.98

75.16 76.63 76.67 77.81 77.08 76.31

574 781 546 521 859 863

36.66 37.26 37.13 36.40 36.92 21.93

83.49 82.94 82.63 83.20 83.58 82.95

78.68 78.40 78.20 79.75 79.22 79.43

SFT-Parallel-7B ALMA-7B-R ALMA-13B-R TowerInstruct-7B-v0.2 TowerInstruct-13B-v0.1 CoT-FT-7B MT-R1-Zero-7B SSR-X-Zero-7B mExTrans-7B

15.77 13.46 14.20 15.54 16.81 15.31 13.75 13.66 10.43

76.73 75.12 75.89 75.49 76.01 76.45 76.93 76.82 76.16

75.37 74.97 75.81 74.93 75.46 74.96 76.72 76.14 76.62

52 69 54 610

40.64 33.75 34.22 38.20 39.26 40.41 34.51 27.01 25.54

84.08 81.12 81.63 83.47 83.85 83.79 82.71 80.08 78.76

79.35 77.72 77.78 78.67 78.44 79.27 79.59 76.57 77.19

TwT-Qwen2.5-7B-Instruct TwT-Qwen2.5-14B-Instruct

17.93 19.10

78.35 78.40

77.30 76.93

281 354

40.12 41.26

84.04 84.17

80.05 79.68

Large Language Models 17.94 74.91 19.70 74.97 17.65 75.04 Large Reasoning Models 593 17.05 74.76 345 19.57 75.03 343 15.52 73.54 403 16.76 75.03 492 18.83 75.49 583 12.36 73.59 MT-Specialized Models 21.09 74.66 14.09 71.92 14.75 72.76 10.99 69.02 11.41 69.23 35 20.52 74.04 56 13.05 72.84 36 13.97 72.86 452 11.80 73.88 Our Models 222 21.13 75.45 234 22.20 75.71

80.33 79.71 80.64

-

41.43 44.39 41.89

84.15 84.49 84.23

81.90 81.76 81.89

-

33.29 33.91 31.83

83.38 83.09 83.00

82.02 81.72 81.80

-

80.63 79.79 80.19 82.02 81.80 81.47

790 677 511 506 751 963

40.69 43.36 39.82 40.44 43.00 33.95

83.94 84.14 83.60 83.88 84.29 83.27

82.39 81.70 81.79 83.29 83.13 82.82

667 415 346 441 531 735

31.33 32.96 29.99 30.93 32.29 26.61

82.95 82.77 82.48 83.10 83.31 82.20

81.91 81.48 81.78 82.70 82.54 81.85

586 734 420 433 660 693

78.39 79.12 79.63 70.13 69.16 78.16 80.66 79.75 80.56

45 71 46 604

43.76 37.97 40.34 46.81 48.55 44.21 27.28 28.86 25.40

84.40 83.04 83.38 84.58 84.99 84.34 83.31 82.12 81.66

82.47 81.36 81.49 80.80 80.93 82.59 82.69 80.98 81.93

46 71 50 565

32.88 25.11 26.83 30.38 31.72 33.07 25.96 25.28 22.13

82.72 80.64 81.50 81.47 81.90 82.44 82.15 81.47 81.55

80.96 79.74 80.35 79.12 79.26 80.66 81.88 80.94 81.78

42 65 48 551

80.19 80.13

269 336

43.74 44.45

84.32 84.42

82.81 82.89

262 287

33.28 34.35

83.48 83.59

82.06 81.91

274 291

Table 18: Detailed metric breakdown for Table 3. We report in-domain translation results across eight domains, averaged over En→Zh, Zh→En, and De→En.

Conversation

Method BLEU

COMET

KIWI

DeepSeek-V3 Gemini-2.0-Flash GPT-4o

36.76 38.05 38.02

87.03 86.90 86.79

81.48 81.37 81.43

DeepSeek-R1 Gemini-2.0-Flash-Thinking OpenAI-o3-mini OpenAI-o1 GPT-5

33.34 37.21 36.66 35.50 35.25

86.47 86.64 86.31 87.10 87.14

81.37 81.24 81.15 82.81 82.82

SFT-Parallel-7B ALMA-7B-R ALMA-13B-R CoT-FT-7B MT-R1-Zero-7B SSR-X-Zero-7B mExTrans-7B

32.20 29.49 31.86 31.58 32.38 30.65 24.53

84.11 84.67 85.48 84.09 85.88 85.32 84.72

80.62 79.96 80.76 80.63 81.50 81.13 81.15

TwT-Qwen2.5-7B-Instruct TwT-Qwen2.5-14B-Instruct

35.10 35.35

86.37 86.39

81.69 81.58

BLEU

COMET

Ecommerce Tokens

BLEU

Large Language Models 32.29 85.53 32.41 85.49 32.88 85.57 Large Reasoning Models 534 27.41 85.03 1204 31.50 85.24 290 31.51 85.13 327 29.65 85.61 448 29.09 85.37 MT-Specialized Models 26.94 82.92 25.13 82.98 26.24 83.75 31 26.71 82.95 53 26.98 84.25 37 25.64 83.94 464 20.46 83.52 Our Models 231 30.92 84.85 240 31.04 85.11

Culture

Method

KIWI

DeepSeek-V3 Gemini-2.0-Flash GPT-4o

40.27 39.11 38.65

85.46 85.02 85.24

83.21 82.94 83.14

DeepSeek-R1 Gemini-2.0-Flash-Thinking OpenAI-o3-mini OpenAI-o1 GPT-5

35.88 36.94 34.05 34.57 35.79

85.45 84.63 84.36 85.06 85.04

83.14 83.68 84.25 84.25 84.15

SFT-Parallel-7B ALMA-7B-R ALMA-13B-R CoT-FT-7B MT-R1-Zero-7B SSR-X-Zero-7B mExTrans-7B

31.81 33.25 22.23 30.41 31.86 28.70 27.12

82.87 83.81 81.12 82.46 83.99 83.25 84.30

81.35 82.83 79.48 81.10 82.85 81.11 83.92

TwT-Qwen2.5-7B-Instruct TwT-Qwen2.5-14B-Instruct

35.24 37.70

84.59 84.71

83.64 83.23

COMET

Social

KIWI

Tokens

BLEU

COMET

KIWI

Tokens

81.19 80.87 81.04

-

32.26 32.98 32.75

84.59 84.28 84.47

81.44 81.04 81.08

-

80.91 80.74 80.97 82.13 82.02

552 822 363 399 609

27.22 31.79 31.79 29.21 29.00

83.94 83.89 84.00 84.50 84.10

81.17 80.83 80.96 82.43 82.31

554 1081 372 405 652

80.12 79.50 80.12 80.11 81.25 80.79 81.24

45 66 50 566

25.88 25.76 27.00 24.98 26.39 25.47 19.60

81.61 82.51 83.06 81.37 83.28 83.19 82.29

80.12 79.77 80.47 79.83 81.40 81.26 81.70

42 65 49 555

81.35 81.21

273 309

30.79 31.04

84.08 83.91

81.58 81.46

269 298

KIWI

Tokens

BLEU

COMET

KIWI

Tokens

79.98 79.38 79.87

-

34.83 34.83 34.97

85.59 85.28 85.47

81.46 81.12 81.31

-

79.99 79.68 80.85 80.95 80.75

602 2335 436 392 530

30.47 33.92 32.45 31.47 31.24

85.08 84.98 84.82 85.45 85.22

81.32 81.23 81.64 82.52 82.41

561 1332 411 413 645

78.98 80.00 78.97 78.88 79.53 79.80 79.23

33 51 34 470

27.81 27.38 26.94 27.21 28.42 26.88 22.06

82.68 83.35 83.16 82.57 84.07 83.70 83.35

80.24 80.41 79.96 80.11 81.31 80.82 81.45

41 63 47 537

80.07 79.98

219 259

32.23 32.99

84.84 84.90

81.67 81.49

269 288

CommonSense Tokens

BLEU

COMET

Large Language Models 32.55 85.36 31.60 84.71 32.54 85.26 Large Reasoning Models 560 28.50 84.52 1220 32.15 84.50 596 28.23 84.28 542 28.43 84.97 984 27.08 84.43 MT-Specialized Models 22.19 81.91 23.25 82.80 27.35 82.41 54 22.38 81.97 79 24.50 82.94 66 23.93 82.82 631 18.61 81.92 Our Models 352 29.08 84.30 333 29.81 84.36

Average

Table 19: Detailed metric breakdown for Table 4. We report OOD translation results across five domains, averaged over En→Zh, Zh→En, and De→En.

En→Zh

Method BLEU

COMET

KIWI

Qwen2.5-7B-Instruct Gemma-2-9B-IT

35.05 32.66

85.44 84.49

83.11 82.36

ALMA-7B-R Tower-Plus-9B SFT-Parallel-7B MT-R1-Zero-7B SSR-X-Zero-7B mExTrans-7B

26.69 37.91 38.58 32.00 34.20 28.63

84.13 86.76 86.34 86.23 86.86 86.44

81.43 84.88 83.39 84.34 84.58 85.02

TwT-Qwen2.5-7B-Instruct TwT-Gemma-2-9B-IT

38.28 36.34

87.27 86.65

84.43 84.23

Zh→En Tokens

BLEU

En→X

Method BLEU

COMET

KIWI

Qwen2.5-7B-Instruct Gemma-2-9B-IT

5.97 14.05

55.56 75.63

51.30 73.16

ALMA-7B-R Tower-Plus-9B SFT-Parallel-7B MT-R1-Zero-7B SSR-X-Zero-7B mExTrans-7B

2.78 6.07 1.19 5.91 6.02 5.09

58.13 63.86 52.16 58.76 58.94 61.33

73.13 69.46 43.97 58.27 56.34 61.74

TwT-Qwen2.5-7B-Instruct TwT-Gemma-2-9B-IT

5.99 11.12

59.57 72.02

57.83 80.82

COMET

Large Language Models 12.63 69.88 13.57 72.41 MT-Specialized Models 13.46 75.12 16.56 77.88 15.77 76.73 62 13.75 76.93 50 14.83 77.45 537 10.43 76.16 Our Models 298 17.93 78.35 227 19.17 78.11

De→En

KIWI

Tokens

BLEU

COMET

KIWI

Tokens

70.22 71.94

-

28.82 25.84

79.36 77.06

80.53 79.01

-

74.97 77.78 75.37 76.72 77.27 76.62

69 54 610

28.87 36.56 35.72 25.56 26.65 22.06

79.66 82.21 81.63 80.39 80.39 79.27

80.13 82.36 81.11 81.73 81.55 80.88

66 45 544

77.30 77.07

281 249

35.53 36.73

81.93 82.15

81.90 81.97

256 218

X→En Tokens

BLEU

COMET

Large Language Models 20.45 76.22 31.55 84.71 MT-Specialized Models 12.01 63.94 27.99 81.38 18.47 75.58 358 21.91 77.48 306 20.73 78.05 1047 15.95 76.80 Our Models 483 22.94 77.73 280 32.79 85.02

Average

KIWI

Tokens

BLEU

COMET

KIWI

Tokens

74.30 82.88

-

20.58 23.53

73.29 78.86

71.89 77.87

-

58.17 79.68 73.05 75.61 75.78 75.21

74 48 731

16.76 25.02 21.95 19.82 20.49 16.43

72.20 78.42 74.49 75.96 76.34 76.00

73.57 78.83 71.38 75.34 75.10 75.89

126 101 694

75.84 83.30

328 257

24.13 27.23

76.97 80.79

75.46 81.48

329 246

Table 20: Detailed metric breakdown for Table 5. We report results on seen and unseen language directions. En, Zh, and De are seen languages, while X denotes unseen languages; En→X and X→En report averages over English↔unseen-language directions.

Related documents

• When Does Reasoning Help in Machine Translation? A Hierarchical Analysis of LRM Reasoning Traces#1035213
• Choosing the Right Language Mode at Inference Time for Multilingual Reliability#927921
• Latent Chain-of-Thought as Planning: Decoupling Reasoning from Verbalization#1118470
• Learning Composable Chains-of-Thought#798613
• Learning Composable Chains-of-Thought#797200