Paper deep dive
Translation with Thought: Difficulty-Adaptive Reasoning via Reinforcement Learning for Multi-Domain Machine Translation
Yongshi Ye, Biao Fu, Chongxuan Huang, Yidong Chen, Xiaodong Shi
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 8/3/2026, 2:53:27 AM
Summary
The paper introduces TwT (Translation with Thought), a resource-rational framework for Multi-Domain Machine Translation (MDMT) that adapts reasoning effort based on input difficulty. It utilizes a two-stage training process: supervised fine-tuning on difficulty-aware chain-of-thought traces distilled from DeepSeek-R1 and refined by GPT-4o, followed by reinforcement learning with a hybrid reward. TwT-7B and TwT-14B outperform larger state-of-the-art reasoning models in translation quality while reducing token usage by 32-60%.
Entities (14)
Relation Signals (10)
TwT → reduces → Token usage
confidence 95% · reducing token usage by 32--60%
TwT → targets → Multi-Domain Machine Translation
confidence 95% · we propose TwT... for Multi-Domain Machine Translation
TwT → uses → Chain-of-Thought
confidence 95% · TwT is trained... on difficulty-aware long chain-of-thought traces
TwT → uses → Reinforcement Learning
confidence 95% · TwT is trained in two stages... (2) reinforcement learning with a hybrid reward
DeepSeek-R1 → generates → Reasoning Traces
confidence 90% · DeepSeek-R1 generates diverse reasoning traces
TwT → outperforms → OpenAI o1
confidence 90% · TwT-7B and TwT-14B outperform much larger SOTA reasoning models... such as... OpenAI-o1
TwT → outperforms → DeepSeek-R1
confidence 90% · TwT-7B and TwT-14B outperform much larger SOTA reasoning models... such as DeepSeek-R1
GPT-4o → rewrites → Reasoning Traces
confidence 90% · GPT-4o... rewriting the traces to match the appropriate reasoning depth
TwT → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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 resource-rational 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-of-thought 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 out-of-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.
Tags
Links
- Source: https://arxiv.org/abs/2607.29287v1
- Canonical: https://arxiv.org/abs/2607.29287v1
Trouble viewing inline? Open PDF directly →
Full Text
116,384 characters extracted from source content.
Expand or collapse full text
Translation with Thought: Difficulty-Adaptive Reasoning via Reinforcement Learning for Multi-Domain Machine Translation Yongshi Ye 1,3 , Biao Fu 2,3, * , Chongxuan Huang 2,3 , Yidong Chen 2,3 , Xiaodong Shi 1,2,3, * 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 Abstract Multi-domain machine translation (MDMT) poses a unique challenge due to varying levels of linguistic complexity across domains. In- spired by human translators’ ability to adapt reasoning effort based on difficulty, we propose TwT(Translation with Thought), a resource- rational framework that learns to modulate in- ference between intuitive and deliberate reason- ing.TwTis trained in two stages: (1) supervised fine-tuning on difficulty-aware long chain-of- thought traces distilled from DeepSeek-R1 and rewritten by GPT-4o to reflect human-like rea- soning economy, and (2) reinforcement learn- ing with a hybrid reward to optimize translation quality and reasoning efficiency. Evaluated on 15 benchmarks spanning in-domain and out- of-domain settings, as well as 3 seen and 59 unseen languages, with ablations across three backbone models,TwT-7BandTwT-14Boutper- form 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) re- mains 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 reason- ing to resolve ambiguity or domain-specific con- structs. However, most MT systems translate all inputs uniformly, lacking mechanisms to adjust inference effort based on domain-specific complex- ity (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 pro- cessing (System 1) for familiar content and slower, deliberate reasoning (System 2) when encounter- ing Rich Points (Agar, 1994), such as ambiguous terminology, complex syntax, or cultural dispari- ties. Because the density of such Rich Points varies across domains and correlates with translation dif- ficulty (Lacruz, 2017), current MT systems still largely lack this adaptive reasoning ability. From the perspective of reasoning allocation, ex- isting 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 large- scale 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 low- resource settings. Recent efforts have introduced Chain-of-Thought (CoT) prompting into transla- tion (Wang et al., 2025a), but this does not fun- damentally 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 DeepSeek- R1 (Guo et al., 2025), represents a shift to the op- posite extreme—an overcommitment to System 2. Reinforcement learning (RL) is often used to train these models to generate Long CoT traces, apply- ing 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 repeti- tive, shallow templates, failing to develop domain- specific reasoning behaviors. The second explores RL with SFT, which fine-tunes on CoT traces be- arXiv:2607.29287v1 [cs.CL] 31 Jul 2026 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 reason- ing, it lacks control over when such reasoning is needed, leading to verbose traces even for simple inputs. This indiscriminate reasoning may help re- veal Rich Points, but often results in overthinking and excessive token usage, reducing efficiency and human alignment. Despite this, most reasoning- based 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 at- tempt to align reasoning effort explicitly with input difficulty. Inspired by human cognitive flexibility, we pro- poseTwT(Translation with Thought), a resource- rational framework for MDMT that learns to allo- cate inference effort based on input difficulty. As shown in Figure 1,TwTdynamically shifts its rea- soning behavior according to input difficulty, using concise reasoning for routine inputs and deeper reasoning for domain-specific challenges. To im- plement this,TwTfollows the RL with SFT pipeline. In the cold-start stage, it performs multi-agent distillation to construct difficulty-adaptive reason- ing traces across domains: a domain-specialized teacher (DeepSeek-R1) generates diverse reason- ing 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 reason- ing and human-like inference modulation, sup- porting 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 transla- tion quality metrics (BLEU and COMET) with a repetition penalty, guiding the model toward effi- cient, domain-adaptive reasoning through outcome- driven learning. We conduct a comprehensive evaluation ofTwT on 15 benchmarks across in-domain and OOD settings, as well as 3 seen and 59 unseen lan- guages, and perform ablation studies on three dif- ferent backbone models to assess generalization across both domain and linguistic axes. Our re- sults demonstrate thatTwTachieves performance competitive with or superior to SOTA LRMs (e.g., DeepSeek-R1, OpenAI-o1) and surpass strong MT- specialized baselines, while reducing token usage by 32–60%. Empirical analysis confirms thatTwT effectively modulates reasoning effort according to task difficulty, leading to more coherent reason- ing processes and more accurate translations. This validates the core intuition behindTwT: 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, do- main 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 hu- man translation workflows, improving both per- formance 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 sig- nals (He et al., 2025), MT-R1-Zero combines rule- based and neural metrics (Feng et al., 2025a), Deep- Trans 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 reason- ing to improve MDMT efficiency and quality. 3 Preliminary Analysis 3.1 Reasoning Collapse Reasoning TemplateRate I will translate this Chinese sentence into En- glish 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 Ap- pendix E.1. This setup is motivated by recent find- ings that RL alone can induce spontaneous reason- ing 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 regu- larization suppresses exploration (Yeo et al., 2025), we monitor token length dynamics both with and without the KL term. However, unlike the “Aha mo- ments” observed in STEM tasks, our experiments reveal rapid mode collapse. As shown in Figures 2(a) and 2(b), reason- ing traces quickly collapse into shallow patterns (≤ 100tokens) regardless of the KL setting. Con- cretely, the model shifts toward high-frequency template recitation, suppressing diverse reasoning. As shown in Table 1, the top three templates ac- count for about 73% of all generated traces in the Zh→En direction. This degeneration reveals a key misalignment: MDMT requires domain-aware rea- soning, yet without proper initialization, the model produces reasoning that is too brief and overly tem- plated to be elicited reliably. To address this, our cold-start phase explicitly initializes adaptive rea- soning behavior before RL. Model EasyHard Quality Token Time Quality Token Time In-Domain SFT-Parallel64.85131261.724015 General-CoT 61.543117256.5355180 Domain-CoT 67.3655813662.65883149 Out-of-Domain SFT-Parallel67.5213759.11368 General-CoT 66.593549660.5551797 Domain-CoT 69.8163412562.52840138 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. 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 (Sys- tem 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 Ta- ble 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 com- petitive on OOD data, particularly on harder sam- ples: when the input does not closely match the do- main patterns seen in training, intermediate reason- ing helps the model better handle syntax, ambigu- ity, and contextual inference than standard parallel SFT. Importantly, adding explicit domain reason- ing 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 re- gardless of input difficulty, resulting in substantial redundancy. On Easy samples, this redundancy is clearly detrimental: despite using 23.9×more to- kens in-domain (311 vs. 13) and 27.2×more on OOD data (354 vs. 13), it still underperforms SFT- 025050075010001250 RL Training Steps 50 100 150 200 250 300 Response Length Response Length vs Training Step 025050075010001250 RL Training Steps 18 21 24 27 30 33 BLEU BLEU vs Training Step 025050075010001250 RL Training Steps 78 79 80 81 82 83 COMET COMET vs Training Step 025050075010001250 RL Training Steps 78 79 80 81 82 83 CometKiwi CometKiwi vs Training Step BLEU COMET CometKiwi BLEU+COMET BLEU+CometKiwi (a) Pure RL training with KL regularization. 025050075010001250 RL Training Steps 50 100 150 200 250 300 Response Length Response Length vs Training Step 025050075010001250 RL Training Steps 22 24 26 28 30 32 BLEU BLEU vs Training Step 025050075010001250 RL Training Steps 78 79 80 81 82 83 COMET COMET vs Training Step 025050075010001250 RL Training Steps 78 79 80 81 82 83 CometKiwi CometKiwi vs Step BLEU COMET CometKiwi BLEU+COMET BLEU+CometKiwi (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, respec- tively. 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 prompt- level domain awareness alone cannot resolve over- thinking, motivatingTwT, which adaptively shifts between System 1 and System 2 to jointly address domain sensitivity and reasoning efficiency. 4 Method We proposeTwT, a resource-rational approach that dynamically allocates reasoning effort by input dif- ficulty, 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 capabili- ties. 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 differ- ent 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 diver- gence, 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 mini- mize token usage, whereas Hard inputs retain com- prehensive 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 ini- tial policy distribution over difficulty-aware reason- ing strategies, enabling the model to autonomously modulate its reasoning depth. Representative exam- ples 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 en- courage the model to identify Rich Points and al- Stage1: Dataset Curation & Cold Start Stage2: RL Training ③ CoT-FT Translate the following src_lang text into tgt_lang while maintaining the domain style of the source text. TwT In-domain test set 퐀᐀퐀䠀퐀退=1 퐀᐀퐀䠀퐀退=−1 퐀᐀퐀䠀퐀退 = 퐀ئج + 퐀 퐀᐀퐀䠀퐀退=−2 퐀 퐀 =퐀 퐀 = 퐀 퐀 , ..., 퐀_퐀 +퐀 −1 퐀᐀퐀䠀퐀退=1− |퐀᐀퐀 퐀 퐀 | |퐀 퐀 | TwT Repetition Penalty Quality RewardFormat Reward ② Difficulty-Aware CoT Data Curation① Domain-Aware CoT Data Curation Estimate the translation difficulty & Optimize the reasoning trace to match its complexity level. DeepSeek-R1 GPT-4o Rewrite Traces Difficulty Evaluation Identify the domain & Generate domain- specific translation reasoning traces. TwT-7BTwT-14B Model Size In-Domain Test Set Comparison Figure 3: Overview ofTwTtraining pipeline.TwTis 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 mit- igating indiscriminate overthinking (Section 3.2), we improve reasoning efficiency and align the model with the resource-rational principle intro- duced in Section 1. The final rewardris crafted from three components to ensure alignment: r = r f + r q − λ· r rep Format Reward (r f ). We employ a binary for- mat reward (r f ∈ 1,−1) to strictly enforce the reasoning-translation structure defined in Figure 5. Hybrid Quality Reward (r q ). To prevent the model from producing plausible but functionally ineffective reasoning, we introduce a verifiable hy- brid feedback signal. Building upon our prelimi- nary 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 seman- tic reward scores, the lexical accuracy (BLEU) de- grades 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 trans- lation fidelity. To remedy this and strictly enforce alignment, we adopt a hybrid quality reward that combines BLEU (B) with COMET (C): r q = ( B(ˆy, y) + C(x, ˆy, y) if r f = 1 −2if r f =−1 This design effectively stabilizes the optimiza- tion process. As evidenced by the training dynam- ics of ourTwTmodels (Figure 8(a)), both the 7B and 14B models (Figures 8(b) and 8(c)) exhibit synchronous improvements in lexical and seman- tic metrics without divergence, validating that the hybrid signal successfully grounds the reasoning process in accurate translation outcomes. N-gram Repetition Penalty (r rep ). We penal- ize redundant reasoning to discourage degenerate loops and promote efficient token usage. For a given reasoning tracec, letG(c)denote the list of contiguousn-grams. In our experiments, we set n = 20. We compute the ratio of repeatedn-grams to discourage repetitive, loop-like patterns: r rep = 1− |set (G(c))| |G(c)| ∈ [0, 1] 5 Experiments 5.1 Experimental Settings Dataset. We use two datasets for training: (1) a curated set of 7K difficulty-adaptive Long CoT ex- amples spanning 10 domains and three translation directions (De→En, En→Zh, Zh→En) for cold- start 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 addi- tionally include multilingual test sets covering both seen and unseen language pairs. Full dataset details are provided in Appendix B. Method LawsNewsScienceSubtitlesLiteraryITKoranMedicalAverage QualityTokensQualityTokensQualityTokensQualityTokensQualityTokensQualityTokensQualityTokensQualityTokensQualityTokens Large Language Models DeepSeek-V377.72-69.41-69.09-62.93-56.83-66.97-57.73-69.16-66.23- Gemini-2.0-Flash76.57-69.30-68.80-62.96-57.41-66.54-58.13-70.21-66.24- GPT-4o73.77-68.61-68.09-62.92-57.47-66.37-57.78-69.34-65.54- Large Reasoning Models DeepSeek-R177.7257768.4749868.4647861.8151453.9557466.2859357.4879069.0166765.40586 Gemini-2.0-Flash-Thinking76.2170268.24114968.17109261.9870857.2378166.2034558.1367769.7341565.74734 OpenAI-o3-mini71.6142868.1844368.0138562.4235556.9754665.9934356.4251168.4034664.75420 OpenAI-o173.8547868.6740868.6036762.6234057.2852166.4540357.9450669.2044165.58433 GPT-576.2878469.0774068.5860662.7451956.2785966.5749258.7175170.1453166.05660 QwQ-32B71.8466767.8858467.8656361.8058455.0986361.4458355.8196366.6873563.55693 MT-Specialized Models SFT-Parallel-7B76.58-66.08-66.38-62.87-55.96-68.02-58.05-70.21-65.52- ALMA-7B-R67.88-63.37-62.77-59.38-54.52-64.20-55.04-67.46-61.83- ALMA-13B-R70.11-64.65-64.23-60.22-55.30-64.54-55.71-68.40-62.89- TowerInstruct-7B-v0.273.91-65.93-65.45-61.07-55.32-66.78-50.05- 70.73-63.66- TowerInstruct-13B-v0.174.65-66.90-66.20-61.89-56.09-67.18-49.93-71.49-64.29- CoT-FT-7B76.725166.244266.073962.762955.575267.823557.574570.384665.3942 MT-R1-Zero-7B68.937267.416466.986161.975555.806965.605655.527164.437163.3365 SSR-X-Zero-7B69.595666.005266.774961.893955.545461.223655.534663.995062.5648 mExTrans-7B70.1159765.4955366.1354659.5247654.4061060.5045255.4160463.0056561.82551 Our Models TwT-Qwen2.5-7B-Instruct75.3531068.4231168.2429463.0324757.8628168.0722258.9226970.2926266.27274 TwT-Qwen2.5-14B-Instruct76.5832068.6228568.3227262.9724158.1435468.3723459.3533670.5928766.62291 Table 3: In-domain translation performance across eight domains, averaged over En→Zh, Zh→En, and De→En. The Bold and underlinedvalues denote the highest and second highest scores, respectively. Method ConversationEcommerceSocialCultureCommonSenseAverage QualityTokensQualityTokensQualityTokensQualityTokensQualityTokensQualityTokens Large Language Models DeepSeek-V368.42-66.34-66.10-69.65-65.96-67.29- Gemini-2.0-Flash68.77-66.26-66.10-69.02-65.23-67.08- GPT-4o68.75-66.50-66.10-69.01-65.89-67.25- Large Reasoning Models DeepSeek-R167.0653464.4555264.1155468.1656064.3460265.62561 Gemini-2.0-Flash-Thinking68.36120465.8382265.50108168.42122065.44233566.711332 OpenAI-o3-mini68.0429065.8736365.5837267.5559664.4543666.30411 OpenAI-o168.4732765.8039965.3840567.9654264.7839266.48413 GPT-568.4044865.4960965.1465268.3398464.0953066.29645 MT-Specialized Models SFT-Parallel-7B65.64-63.33-62.54-65.34-61.03-63.58- ALMA-7B-R64.71-62.54-62.68-66.63-62.02-63.71- ALMA-13B-R66.03-63.37-63.51-60.94-62.91-63.35- CoT-FT-7B65.433163.264562.064264.665461.083363.3041 MT-R1-Zero-7B66.595364.166663.696566.237962.325164.6063 SSR-X-Zero-7B65.703763.465063.314964.356662.183463.8047 mExTrans-7B63.4746461.7456661.2055565.1163159.9247062.29537 Our Models TwT-Qwen2.5-7B-Instruct67.7223165.7127365.4826967.8235264.4821966.25269 TwT-Qwen2.5-14B-Instruct67.7724065.7930965.4729868.5533364.7225966.46288 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-Factory 1 (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 of1e−5, a total batch size of 32, a cosine learning rate scheduler, a warm-up ra- tio of 0.1, a maximum input sequence length of 4096, and DeepSpeed ZeRO Stage 3. The cold- start stage completes within 10 minutes. For RL, we useverl 2 (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 rate1e−6, KL loss coef- ficientβ = 1e−3, maximum response length 2048, and repetition penalty withn = 20. RL training 1 https://github.com/hiyouga/LLaMA-Factory 2 https://github.com/volcengine/verl takes about 10 hours. During inference, we use vLLM 3 (Kwon et al., 2023) for efficient decoding with temperature 0.0 and repetition penalty 1.05. Metrics. We report Quality, defined as the av- erage of BLEU, COMET (Rei et al., 2020), and CometKiwi (Rei et al., 2022), and Tokens, the av- erage 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), GPT- 4o (OpenAI et al., 2024), and open-source mod- els 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), Gemini- 3 https://github.com/vllm-project/vllm Method En→ZhZh→EnDe→EnEn→X→EnAverage QualityTokensQualityTokensQualityTokensQualityTokensQualityTokensQualityTokens Large Language Models Qwen2.5-7B-Instruct67.87-50.91-62.90-37.61-56.99-55.25- Gemma-2-9B-IT66.50-52.64-60.64-54.28-66.38-60.09- MT-Specialized Models ALMA-7B-R64.08-54.52-62.89-44.68-44.71-54.18- Tower-Plus-9B 69.85-57.41-67.04-46.46-63.02-60.76- SFT-Parallel-7B69.44-55.96-66.15-32.44-55.70-55.94- MT-R1-Zero-7B67.526255.806962.566640.9835858.337457.04126 SSR-X-Zero-7B68.555056.525462.864540.4330658.194857.31101 mExTrans-7B66.7053754.4061060.7454442.72104755.9973156.11694 Our Models TwT-Qwen2.5-7B-Instruct69.9929857.8628166.4525641.1348358.8432858.85329 TwT-Gemma-2-9B-IT69.0722758.1224966.9521854.6528067.0425763.17246 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 In-DomainOut-of-Domain QualityTokensQualityTokens TwT-Qwen2.5-7B-Instruct64.7727866.28263 w/o RP64.5230166.12271 w/o RP + w/o Adaptive CoT64.6874866.17720 w/o RP + w/o Cold Start64.186265.2959 SFT only w/ Adaptive CoT62.4725364.09219 SFT only w/ domain-aware CoT61.6547963.47452 SFT only w/ general CoT61.3953163.11519 Backbone Models Qwen2.5-7B-Instruct61.99-64.71- Llama-3.1-8B-Instruct59.84-63.33- Gemma-2-9B-IT59.93-63.68- Our Models TwT-Qwen2.5-7B-Instruct64.7727866.28263 TwT-Llama-3.1-8B-Instruct63.2030564.44298 TwT-Gemma-2-9B-IT64.7123166.05220 Table 6: Ablation study on in-domain and OOD transla- tion 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 (Qwen- Team, 2025); (3) MT-specialized models, including non-reasoning LLMs such as TowerInstruct (Alves et al., 2024), ALMA-R (Xu et al., 2024a,b), SFT- Parallel (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 reasoning- oriented models including MT-R1-Zero (Feng et al., 2025a), mExTrans (Wang et al., 2025c), and SSR- X-Zero (Yang et al., 2025). For fairer comparison, Appendix C details the training data size of MT- specialized 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 mod- els including GPT-5 (66.05). Notably, our smaller variant,TwT-7B, also surpasses dedicated MT sys- tems, demonstrating that our method scales effec- tively with model size. Compared to mExTrans- 7B, which is also trained under the R1 paradigm, TwT-7Byields 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 struc- turally complex and culturally nuanced domains. In the Literary domain,TwT-14Bconsistently out- performs 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-R1- Zero-7B by+2.34 points. These results suggest that, for multi-domain translation, neither shallow System 1 execution, indiscriminate System 2 over- thinking, 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-7Bachieves a strong average score of 66.25 on five OOD test sets, surpassing MT- specialized baselines such as SFT-Parallel-7B (63.58), which exhibit poor generalization un- der domain shift. Moreover,TwT-7Bexhibits 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 com- pact, internalized translation procedures. These results indicate thatTwTcaptures domain-agnostic translation logic rather than relying on domain- specific memorization, approaching the SOTA per- formance of GPT-4o at a fraction of the computa- tional cost and model size. 5.3 Multilingual Generalization On seen directions,TwT-7Bimproves 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 exam- ples, outperforms the multilingual Tower-Plus-9B, built on the same backbone but trained on 286K examples, by a substantial margin of+8.19. Over- all,TwTachieves the highest average score across all directions, indicating that its reasoning mecha- nism generalizes beyond language boundaries and captures transferable alignment strategies. 5.4 Ablation Study We conduct a comprehensive ablation study to as- sess the contribution of each component in theTwT framework and validate its generalizability across backbone architectures (Table 6). Removing the repetition penalty (w/o RP) leads to a slight qual- ity 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 adap- tive rewriting in controlling verbosity and ensuring inference efficiency. We further examine the neces- sity of the two-stage training pipeline. Eliminating the cold-start SFT phase (w/o RP + w/o Cold Start) causes performance degradation and length col- lapse (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 re- sult (62.47), followed by domain-aware (61.65) and general CoT (61.39), showing a clear performance hierarchy. Still, only the fullTwTpipeline 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 ap- plyTwTto three backbone models: Qwen2.5-7B, 4 En→X and X→En are averaged over 59 unseen lan- guages from FLORES+; see Appendix B.6 for the full list. Llama-3.1-8B, and Gemma-2-9B, and observe con- sistent improvements. For instance,TwTimproves Gemma-2-9B-IT from 59.93 to 64.71 in-domain and from 63.68 to 66.05 OOD, demonstrating that TwTis a model-agnostic framework that robustly enhances translation reasoning regardless of the underlying architecture. 6 Empirical Analysis 6.1 Human Reasoning Alignment To benchmarkTwT’s reasoning against human cog- nition, we conduct a qualitative analysis on 10 Zh→En examples, with expert commentary from a translation studies faculty member. A representa- tive case is shown in Appendix G.2. Cognitive Convergence. The analysis revealed thatTwT’s reasoning exhibits strong parallels with human translators in early-stage decision-making. Specifically,TwTeffectively (1) identifies transla- tion domain and stylistic register, (2) handles com- plex sentence structures with appropriate syntac- tic parsing, and (3) demonstrates context-aware terminology adaptation. For example, it consis- tently distinguishes between literary and techni- cal expressions and adjusts lexical choices accord- ingly. Its structured CoT mirrors key aspects of professional reasoning—such as coherence main- tenance, discourse flow control, and sensitivity to stylistic norms—indicating thatTwThas internal- ized domain-aware reasoning behavior resembling human translation logic. Pragmatic Divergence.Despite these strengths, TwTstill shows gaps compared with expert transla- tors. It occasionally struggles with cross-sentence consistency in terminology, especially when han- dling long-form repetitions or abbreviated refer- ences. Moreover, its output lacks fine-grained con- trol over tone, idiomaticity, and cultural adapta- tion, which human translators adjust based on prag- matic context and target audience. These issues suggest thatTwT’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 process- oriented feedback, thereby fostering deeper prag- matic alignment with human cognitive processes. 6.2 Reasoning Redundancy Reduction To assess whetherTwTeliminates unnecessary com- putation, we employ DeepSeek-V3.2 to detect six Redundancy TypeResolved (%)Not Resolved (%) Over-segmentation87.312.7 Unnecessary linguistic explanation93.26.8 Semantic repetition96.93.1 Irrelevant information93.76.3 Redundant alternative translations96.53.5 Low-density long descriptions97.22.8 Table 7: Resolution rates of six redundancies by TwT. Method EasyMediumHardAll QualityTokensQualityTokensQualityTokensQualityTokens DeepSeek-R166.3048063.8255662.1557964.09538 Gemini-2.0-Flash-Thinking66.6243264.3475462.93103564.63740 OpenAI-o3-mini66.5728063.4443061.4459663.82435 OpenAI-o167.1430264.2643262.4957864.63437 QwQ-32B62.9157761.9275860.5284461.78726 General-CoT61.5431160.2743356.5355159.45432 TwT-Qwen2.5-7B-Instruct67.9321664.8527262.7833265.19273 TwT-Qwen2.5-14B-Instruct67.9820765.2330063.2637865.49295 Table 8: In-domain performance by difficulty level. distinct forms of reasoning redundancy across 15 domains, utilizing the prompt provided in Fig- ure 12. As shown in Table 7,TwTdemonstrates exceptional efficiency, successfully resolving over 94% of the redundant steps observed in the SFT- with-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 adap- tive reasoning behavior, effectively activating Sys- tem 2 reasoning for Rich Points while avoiding un- necessary elaboration on straightforward segments. 6.3 Translation Difficulty Adaptation To better understand how models adapt their rea- soning 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, av- eraged across all domains. Results indicate that TwTeffectively addresses the reasoning redundancy of General-CoT through resource-rational alloca- tion. On Easy inputs, it reduces token usage by 33% while improving quality by+6.4 points; conversely, on Hard inputs, it focuses on performance, achiev- ing a substantial+6.7 points quality gain. More- over,TwTmodels 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% com- pared to OpenAI-o3-mini and by 60% compared to Gemini-2.0-Flash-Thinking. This highlightsTwT’s ability to generate concise, difficulty-aware reason- 0.40.20.00.20.4 Principal Component 1 0.6 0.4 0.2 0.0 0.2 0.4 Principal Component 2 Semantic Space Exploration (PCA) Pure-RL TwT (Ours) 0.00.20.40.60.81.0 Cosine Similarity 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Density Pairwise Reasoning Similarity Distribution Pure-RL TwT (Ours) 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 degen- erates into shallow, repetitive templates. To verify whetherTwTsuccessfully mitigates this reasoning collapse, we conduct a semantic diversity analy- sis across 15 domains. We compute pairwise co- sine similarities of generated reasoning traces using multilingual Sentence-BERT 5 . Pure RL exhibits se- vere redundancy with a mean similarity of 0.89, whereasTwTsignificantly 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 com- pared to the broad semantic manifold ofTwT, and the similarity histogram (Right) confirms thatTwT diffuses the sharp redundancy peak of the baseline into a balanced distribution. These results show thatTwTovercomes template dependency and en- courages genuine reasoning. 6.5 Further Analysis Appendix F provides additional analyses ofTwT, including MQM error types, KL ablation, train- ing dynamics, domain-aware prompting, inference cost, language consistency, and failure cases. 7 Conclusion In this work, we presentTwT, a resource-rational translation model that adapts reasoning effort to in- put difficulty.TwTcombines difficulty-aware SFT and hybrid-reward RL to balance System 1 and Sys- tem 2 behavior. Evaluated across diverse domains and languages,TwTmatches or surpasses SOTA LRMs while reducing token usage by 32–60%, vali- dating the effectiveness of aligning translation with human reasoning economy. 5 sentence-transformers/paraphrase-multilingual- MiniLM-L12-v2 Limitations WhileTwTachieves robust performance across mul- tiple domains, several limitations remain. First, the RL training data is randomly sampled without con- trolling 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 inher- ent in those models. Although our current setup yields consistent improvements, such biases could influence the reasoning behavior or stylistic tenden- cies ofTwT. Finally, our current reward design does not incorporate difficulty-aware reward shaping. In particular, no length-based reward is applied to en- courage concise reasoning on simple inputs and more detailed analysis for complex ones. Incorpo- rating such adaptive rewards may further enhance the model’s ability to adjust reasoning depth based on input complexity in MDMT. We leave this di- rection for future work. Acknowledgment This work is supported by the National Sci- ence 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 Five- Year Plan (Grant No. WT135-38), the public tech- nology service platform project of Xiamen City (No. 3502Z20231043). In addition, we used a large language model to assist in polishing the visualiza- tions 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 Asso- ciation for Computational Linguistics, pages 7747– 7763, Online. Association for Computational Lin- guistics. Duarte M. Alves, José Pombal, Nuno M. Guerreiro, Pe- dro H. Martins, João Alves, Amin Farajian, Ben Pe- ters, 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 reason- ing 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, Bingx- uan 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: Advanc- ing 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, Al- buquerque, New Mexico. Association for Computa- tional 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 trans- lation and cognition, pages 263–278. Routledge. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schel- ten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mi- tra, 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 capa- bility 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 rea- soning in neural machine translation. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3662–3672, Online. Association for Computational Linguistics. 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 incen- tivizing translation capability in llms via reasoning learning. Preprint, arXiv:2502.19735. 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 Asso- ciation for Computational Linguistics: EMNLP 2024, pages 5726–5746, Miami, Florida, USA. Association for Computational Linguistics. Kristian Tangsgaard Hvelplund. 2011. Allocation of cognitive resources in translation: An eye-tracking and key-logging study. Frederiksberg: Copenhagen Business School (CBS). Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richard- son, Ahmed El-Kishky, Aiden Low, Alec Hel- yar, 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, Ond ˇ rej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Thamme Gowda, Yvette Graham, Roman Grund- kiewicz, Barry Haddow, Rebecca Knowles, Philipp Koehn, Christof Monz, Makoto Morishita, Masaaki Nagata, Toshiaki Nakazawa, Michal Novák, Martin Popel, and Maja Popovi ́ c. 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 Computa- tional Linguistics. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gon- zalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serv- ing 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. Isabel Lacruz. 2017. Cognitive effort in translation, edit- ing, and post-editing. The handbook of translation and cognition, pages 386–401. Zihao Li, Shaoxiong Ji, and Jörg Tiedemann. 2025. Test- time scaling of reasoning models for machine trans- lation. 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 trans- lation with large reasoning models.Preprint, arXiv:2503.10351. NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Hef- fernan, 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. OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander M ̨adry, 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. OpenAI. 2025a. Introducing GPT-5.https://openai. com/zh-Hans-CN/index/introducing-gpt-5/. OpenAI. 2025b. Openai o3-mini.https://openai. com/index/openai-o3-mini/. Accessed: 2025- 04-21. Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic evalu- ation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Compu- tational 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 Process- ing (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 sub- mission 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 Com- putational Linguistics. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017.Prox- imal 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 mathemati- cal 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 Sys- tems, 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, Yim- ing Wang, and Longyue Wang. 2014. UM-corpus: A large English-Chinese parallel corpus for statistical machine translation. In Proceedings of the Ninth In- ternational 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 learn- ing. Preprint, arXiv:2504.10187. Jiaan Wang, Fandong Meng, and Jie Zhou. 2025c. Ex- trans: 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. Find- ings of the WMT 2024 shared task on discourse-level literary translation. In Proceedings of the Ninth Con- ference on Machine Translation, pages 699–700, Mi- ami, Florida, USA. Association for Computational Linguistics. Longyue Wang, Zhaopeng Tu, Yan Gu, Siyou Liu, Dian Yu, Qingsong Ma, Chenyang Lyu, Liting Zhou, Chao- Hong 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 trans- lation: 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: Teach- ing large language models to translate through self- reflection. In Proceedings of the 62nd Annual Meet- ing of the Association for Computational Linguis- tics (Volume 1: Long Papers), pages 6144–6158, Bangkok, Thailand. Association for Computational Linguistics. Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Has- san 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 Mur- ray, and Young Jin Kim. 2024b. Contrastive prefer- ence 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, Dayi- heng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Ji- axi 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 self- rewarding reinforcement learning for machine trans- lation. 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 Associ- ation for Computational Linguistics: EMNLP 2024, pages 13078–13096, Miami, Florida, USA. Associa- tion 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 As- sociation 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 queryx, the model samples a group ofGresponses y i G i=1 , each scored with rewardr i . The normal- ized advantage for each sample is computed as: A i = r i − mean(r G j=1 ) std(r G j=1 ) .(1) Then GRPO optimizes the policy modelπ θ by max- imizing the following objective: J GRPO (θ) =E x∼D,y i G i=1 ∼π θ old (·|x) 1 G G X i=1 min π θ (y i | x) π θ old (y i | x) A i , clip π θ (y i | x) π θ old (y i | x) , 1− ε, 1 + ε A i − βD KL π θ π ref ! , (2) whereπ θ old andπ θ are the old and current policies, εis the PPO clipping threshold, andβcontrols the weight of the KL regularization. B Datasets B.1 Details of Cold Start Data 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: Template 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 pro- cess in the mind, and then provides the final translation. The translation reasoning pro- cess and the final translation are enclosed within <think> </think> and <answer> </an- swer> tags, respectively, i.e., <think> trans- lation reasoning process here </think> <an- swer> final translation here </answer>. User: Translation question. Assistant: <think> Figure 5: Template for pure RL in MT task. •TheGerman-Englishmulti-domain dataset (Aharoni and Goldberg, 2020), including five distinct domains: IT, Law, Medical, Koran, and Subtitles. •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. For each domain, we randomly select 2K sentence pairs with a minimum source sentence length of 20 words (characters for Chinese) to ensure meaning- ful reasoning potential. This results in a total of 20K training samples used for RL training. B.3 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 thevalid_1, valid_2,test_1, andtest_2subsets to form a comprehensive test set. The data statistics is illus- trated in Table 9. B.4 Out-of-Domain Test Data For out-of-domain evaluation, we consider a di- verse set of test sets spanning multiple language pairs and domains. Specifically, the Conversa- tion, Ecommerce, and Social domains are drawn from the WMT22 shared tasks (Kocmi et al., GuoFeng-Webnovel Literary 684 UM-corpus En-Zh Laws 694 UM-corpus En-Zh News 761 UM-corpus En-Zh Science 637 UM-corpus En-Zh Subtitles 763 Multi-domain De-En IT 768 Multi-domain De-En Koran 761 Multi-domain De-En Law 751 Multi-domain De-En Medical 776 Multi-domain De-En Subtitles 787 Figure 6: Distribution of our curated difficulty-adaptive Long CoT data. DomainNumDomainNum En→Zh Laws456De→En IT2000 En→Zh Subtitles597De→En Koran2000 En→Zh Science503De→En Medical2000 En→Zh News1500De→En Law2000 Zh→En Literary3038De→En Subtitles2000 Table 9: In-domain test sets and the number of samples for En↔Zh and De→En translation tasks. DomainNumDomainNum En→Zh Conversation484En→Zh Social511 En→Zh Ecommerce530En→Zh Culture778 Zh→En CommonSense1200Zh→En Conversation349 Zh→En Social491Zh→En Ecommerce518 De→En Conversation462De→En Social515 De→En Ecommerce501 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 Com- monSense 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 ofTwTacross diverse translation scenarios, we designed the dataset to reflect a broad range of linguistic and contextual com- plexity, 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 de- manding content (e.g., Literature); noisier or in- formal genres (e.g., Subtitles, Conversation, Social media). This diversity ensures comprehensive eval- uation across domain specificity, reasoning diffi- culty, and stylistic variation. B.6 Multilingual Test Data For unseen-language evaluation, we adopt the FLO- RES+ benchmark (NLLB Team et al., 2024) and construct an unseen language set to minimize eval- uation 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), Ko- rean (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 re- flect generalization to genuinely unseen languages rather than memorization of language-specific post- training signals. Next, to enable consistent com- putation of COMET and CometKiwi across all un- seen directions, we further restrict the remaining FLORES+ languages to those supported by our COMET/CometKiwi scorers. After these two fil- ters, we obtain a final unseen-language set consist- ing of 59 languages (Table 12). The seen languages—German (de), English (en), and Chinese (zh)—are used across all base- line training sets, and are covered by the follow- ModelAccuracy (↓)Style (↓)Fluency (↓)Terminology (↓)Non-translation (↓) DeepSeek-V350.5530.506.9511.900.09 Gemini-2.0-Flash52.5529.556.9110.810.18 GPT-4o50.5230.767.1111.400.21 DeepSeek-R150.1330.976.7112.100.09 Gemini-2.0-Flash-Thinking52.3129.706.5411.310.13 OpenAI-o3-mini49.8230.707.1312.230.11 OpenAI-o150.6829.697.1212.370.14 TwT-Qwen2.5-14B-Instruct55.0226.376.8911.640.08 Table 11: MQM-based error analysis across different LLMs and LRMs. odeLanguageCodeLanguageCodeLanguage afrAfrikaansalsAlbanianamhAmharic asmAssamesebelBelarusianbenBengali bosBosnianbulBulgariancatCatalan cymWelshekkEstonianellGreek epoEsperantoeusBasquefilFilipino gleIrishglgGaliciangujGujarati hauHausahebHebrewhrvCroatian hyeArmenianindIndonesianjavJavanese kanKannadakatGeorgiankazKazakh khkMongoliankhmKhmerkirKyrgyz laoLaolitLithuanianlvsLatvian malMalayalammarMarathimkdMacedonian myaBurmesenpiNepalipanPunjabi pbtPashtopltMalagasysanSanskrit sinSinhalaslkSlovakslvSlovenian somSomalisrpSerbiansunSundanese swhSwahilitamTamiltelTelugu thaThaiturTurkishuigUyghur urdUrduvieVietnamesexhoXhosa yddYiddishzsmMalay Table 12: The 59 unseen languagesL unseen used for En↔X evaluation after filtering FLORES+ by (i) post- training language coverage of evaluated backbones and (i) COMET/CometKiwi language support. ing datasets: the German-English multi-domain dataset (Aharoni and Goldberg, 2020), the English- Chinese 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 train- ing data size used by each MT-specialized base- line. WhileTwTand most reasoning-augmented models, including MT-R1-Zero-7B, CoT-FT-7B, SFT-Parallel, and mExTrans-7B, are trained on ap- proximately 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-2.0- flash-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 sam- pling 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 to- ward producing translation-relevant reasoning be- havior, we modify the template used in DeepSeek- R1-Zero (Guo et al., 2025) to better fit the transla- tion 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 rewardr f is set to 1 if the format is correct, and -1 otherwise. Quality Reward: We experiment with several qual- ity metrics to evaluate the final translation out- put. Specifically, we consider BLEU 6 (Papineni 6 https://github.com/mjpost/sacrebleu et al., 2002), a metric based on n-gram lexical overlap, and COMET 7 (Rei et al., 2020) and CometKiwi 8 (Rei et al., 2022), neural metrics that estimate semantic similarity using contextual em- beddings through reference-based or reference-free scoring. All metric scores are normalized to the range[0, 1]before being used in the reward func- tion. During training, the quality rewardr q is com- puted based on one or more of these metrics. It is only applied when the output format is correct; otherwise, we assign a fixed penalty ofr q = −2. The final reward used during RL training is the sum of the format and quality components:r = r f +r q . RL Optimization. We train the model using GRPO algorithm (Shao et al., 2024), which im- proves 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 dif- ferent reward signals, we experiment with five qual- ity 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 Qwen2.5-7B- Instruct (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, us- ing 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 Unbabel/wmt22-comet-da 8 Unbabel/wmt22-cometkiwi-da βBLEUCOMETKIWIQualityLens. 030.0782.3581.2064.54130 0.00130.5882.5281.2164.77278 0.00530.7982.4881.2864.85273 0.0130.3782.4981.4064.75282 0.0230.2482.3781.2964.63254 0.0529.7182.2381.2464.39275 Table 13: In-domain results with different KL coeffi- cient. F Further Analysis F.1 Error-Type Analysis under the MQM Framework To further investigate what types of translation er- rors reasoning helps address, we conducted a de- tailed analysis based on the MQM (Multidimen- sional Quality Metrics) framework. For error clas- sification, we used an external LLM as an annota- tor, excluding "Source Error" as it pertains to low- quality source data. Given the distinct challenges of MDMT compared to general MT, we extended error types such as "Style Inconsistency", "Cross- domain 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 be- lieve 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 fine- tuning. Notably, after incorporating domain-aware CoT data,TwTreduced style errors by 3.18%, sig- nificantly 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 re- moving KL regularization can enhance Long CoT reasoning. To examine whether this holds in the MDMT setting, we investigate the effect of the KL coefficientβin Table 13. However, our findings diverge from this conclusion: settingβto 0—com- pletely removing the KL constraint—results in sig- nificantly shorter and less informative outputs. In contrast, small non-zero values (e.g.,β = 0.001or 0.005) achieve the best overall quality scores while maintaining reasonable response lengths. Larger values (e.g.,β = 0.02or0.05) overly constrain the 025050075010001250 RL Training Steps 1.9 2.0 2.1 2.2 Reward Reward vs Training Step TwT-7B TwT-14B 025050075010001250 RL Training Steps 0.00 0.25 0.50 0.75 1.00 1.25 1.50 Format Error Rate (%) Format Error Rate vs Training Step TwT-7B TwT-14B Figure 7: Average reward and format error rate over RL training steps. policy, slightly degrading translation quality and in- creasing length instability. These results reflect the importance of balancing generation stability and flexibility: moderate KL regularization helps sup- press degenerate behavior while preserving adap- tive, difficulty-aware reasoning traces. F.3 Analysis of Training Dynamics. To better understand how RL shapes TwT’s reason- ing and translation behavior, we analyze the train- ing dynamics ofTwT-7BandTwT-14Bfrom two complementary perspectives. Analysis of Reward and Format Stability Fig- ure 7 further illustrates the reinforcement learning dynamics ofTwT-7BandTwT-14Bin terms of av- erage reward and format error rate. Both models show a rapid increase in reward within the first few hundred steps, followed by a stable plateau, indicat- ing that the hybrid reward formulation enables effi- cient convergence toward high-quality translation reasoning. The largerTwT-14Bmodel consistently achieves higher rewards throughout training, sug- gesting stronger optimization capacity and better utilization of the reward signal. In contrast, the for- mat error rate decreases sharply during the initial phase and remains near zero thereafter, demon- strating that the format reward effectively reduces structural inconsistencies and output anomalies as training progresses. Translation Quality and Reasoning DepthFig- ure 8 shows the training dynamics ofTwT-7B andTwT-14Bduring 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 rea- soning, but decreases in later steps as the model learns to maintain quality with more concise traces. This indicates convergence to efficient, difficulty- Table 14: General prompt vs. domain-aware prompt across domains. Prompt ITLawMedical BLEUCOMETKIWIBLEUCOMETKIWIBLEUCOMETKIWI General prompt36.6683.4978.6839.1684.9782.3540.2383.8281.81 Domain-aware prompt36.5583.5779.8240.5785.3283.5441.1484.0682.97 Table 15: SFT trained on general CoT vs. domain-aware CoT. Method In-DomainOut-of-Domain BLEUCOMETKIWIBLEUCOMETKIWI SFT w/ general CoT23.5180.5280.1525.1283.4080.81 SFT w/ domain-aware CoT23.7480.7480.4625.6783.5781.18 ∆+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 ofTwT-7BandTwT-14Bunder 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 rea- soning depth. Notably,TwT-14Bgenerates shorter responses thanTwT-7Bon easy inputs, but longer responses on medium and hard inputs. This sug- gests 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 experi- ments that evaluate its impact from both the infer- ence and training perspectives. First, we investigate how prompt formulation affects translation quality when the model re- mains fixed (DeepSeek-R1). Two prompts are compared: a general prompt ("Translate the fol- lowing 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 represen- tative domains (IT, Law, Medical) demonstrates that domain-aware prompt explicitly instructs the model to infer and preserve domain-specific stylis- tic and terminological features. This better acti- vates the reasoning capacity and leads to more accurate domain-aware translation. These find- ings motivated us to adopt domain-aware prompts throughoutTwT’s training data curation phase, al- lowing domain signals to be explicitly injected dur- 025050075010001250 RL Training Steps 200 250 300 350 400 450 Response Length Response Length vs Training Step 025050075010001250 RL Training Steps 26.5 27.7 28.9 30.1 31.3 32.5 BLEU BLEU vs Training Step 025050075010001250 RL Training Steps 80.5 81.0 81.5 82.0 82.5 83.0 COMET COMET vs Training Step 025050075010001250 RL Training Steps 80.3 80.5 80.7 80.9 81.1 81.3 CometKiwi CometKiwi vs Training Step TwT-7B TwT-14B (a) Training dynamics for TwT-7B and TwT-14B. 025050075010001250 RL Training Steps 125 175 225 275 325 375 Response Length Response Length vs Training Step 025050075010001250 RL Training Steps 24.0 26.2 28.4 30.6 32.8 35.0 BLEU BLEU vs Training Step 025050075010001250 RL Training Steps 78.0 79.8 81.6 83.4 85.2 87.0 COMET COMET vs Training Step 025050075010001250 RL Training Steps 78.0 79.2 80.4 81.6 82.8 84.0 CometKiwi CometKiwi vs Training Step Easy Medium Hard (b) Training dynamics for TwT-7B under different difficulty level. 025050075010001250 RL Training Steps 125 220 315 410 505 600 Response Length Response Length vs Training Step 025050075010001250 RL Training Steps 24.0 26.2 28.4 30.6 32.8 35.0 BLEU BLEU vs Training Step 025050075010001250 RL Training Steps 78.0 79.8 81.6 83.4 85.2 87.0 COMET COMET vs Training Step 025050075010001250 RL Training Steps 78.0 79.2 80.4 81.6 82.8 84.0 CometKiwi CometKiwi vs Training Step Easy Medium Hard (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 differ- ent prompts to construct Long-CoT data for SFT. Specifically, one dataset is generated using the gen- eral 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 auto- matic 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- 020406080100120 Time Cost (ms/example) TwT-14B (ours) TwT-7B (ours) Qwen2.5-7B-Instruct + SFT 121 57 12 Figure 9: Inference time comparison. ferent models. As shown in Figure 9, the reasoning- augmentedTwTmodels incur additional computa- tional overhead compared with the non-reasoning baseline (Qwen2.5-7B-Instruct + SFT). Specifi- cally,TwT-7BandTwT-14Brequire 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, rea- soning accuracy, and style adaptation demonstrated in previous sections. F.6 Language Consistency Discussion To verify thatTwT’s reasoning traces are linguisti- cally coherent and aligned with the target transla- tion language, we conducted a systematic analysis of language consistency throughout the data cura- tion 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 con- tained incorrect languages. After GPT-4o rewrit- ing, 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 ran- domly sampled instances. Results demonstrate robust consistency across three dimensions: (1) Internal Stability: All models exhibited high self- agreement across 5 independent trials (GPT-4o: 0.92, DeepSeek: 0.93, Gemini: 0.91). (2) Cross- Model Agreement: The three models reached a 90% consensus rate after majority voting. (3) Hu- man 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 pro- vides a reliable and stable proxy for human judg- ment in our difficulty-aware framework. F.8 Structured Analysis of Failure Cases To provide a more systematic analysis, we fur- ther examineTwT’s bad cases across all domains. Specifically, we select the 10 lowest-quality ex- amples 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 de- cision in its reasoning, but the final output fails to realize it. That is, the reasoning identifies an appro- Human EvaluationTwT WinTwT LoseTie TwT-Qwen2.5-14B-Instruct vs. DeepSeek-R1 (671B)0.410.470.12 TwT-Qwen2.5-14B-Instruct vs. DeepSeek-V3 (671B)0.2750.4050.32 TwT-Qwen2.5-14B-Instruct vs. TowerInstruct-13B-v0.10.540.300.16 Table 16: Summary of human evaluation results for TwT compared with other models. priate lexical choice or phrasing, while the predic- tion 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 with- out 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 comple- ment our automatic metrics. Specifically, we ran- domly 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 fromTwT-14Band one from a comparison model—were indepen- dently evaluated by three professional translators, who selected the better translation or marked a tie when the quality difference was negligible. Ta- ble 16 summarizes the human evaluation results across three baselines: DeepSeek-R1, DeepSeek- V3, and TowerInstruct-13B-v0.1. These results show that whileTwT-14Bslightly underperforms DeepSeek-R1 and DeepSeek-V3—expected given its much smaller size—it still achieves a strong de- gree of parity. Importantly,TwT-14Boutperforms TowerInstruct-13B-v0.1 in more than half of the evaluated examples (54% win rate), supporting the effectiveness of its reasoning-driven design in trans- lation. G.2 Comparison with Human Translation Reasoning To further evaluate the alignment betweenTwT’s reasoning process and that of professional human translators, we conducted a case study using a com- plex narrative sentence from a fictional novel, as shown in Figure 17. The source sentence features multiple subordinate and concessive clauses, com- plex logical progression, and dense world-specific terminology—a typical example in the fictional novel domain. Such sentences require precise con- trol of syntax, consistent terminology management, and sensitivity to logical flow and pragmatic tone. Similarities.TwTdemonstrates several key rea- soning behaviors consistent with professional trans- lators. 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 syntac- tic and logical analysis stage in human translation reasoning, reflectingTwT’s strong ability to detect hierarchical structure and causal relationships. Sec- ond,TwTshows strong terminological consistency, accurately translating key domain-specific entities such as "Immortal Holy Water" (不老圣水), "Di- vine Realm Major World" (神属大世界), and "War Star Federation" (战星联邦). While minor stylis- tic improvements remain possible, the model’s use of consistent and semantically clear terms aligns with the terminology standardization step in pro- fessional translation workflows.TwTalso handled subtle pragmatic expressions such as "有交情" and "换取" with contextually appropriate equiva- lents ("have connections with", "trade for"), indicat- ing an emerging awareness of functional-pragmatic equivalence—a key element in expressing commu- nicative intent and contextual tone. Moreover, be- fore translating,TwTdemonstrated an initial assess- ment of text difficulty and style, recognizing that the sentence belongs to a fictional novel with spe- cialized terms. This awareness parallels a human translator’s pre-translation difficulty assessment and background analysis, where genre and register are evaluated to inform strategy.TwTchose to pre- serve the source-style transliteration rather than do- mesticate the names, maintaining consistency with the fantasy universe’s internal logic—an appropri- ate decision for this genre. At the structural level, TwTeffectively 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 syntac- tic and logical relations during target-language reor- ganization, consistent with human translators’ rea- soning in structural decomposition and coherence reconstruction. In lexical judgment,TwTmade 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" ap- propriately 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 pro- cess. Finally,TwT’s output features a natural syn- tactic flow and coherent discourse structure. For instance, "some powerful individuals...had connec- tions 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. Pro- fessional translators would typically balance ter- minological consistency with referential economy, adopting simplified references such as "the Divine Realm" or "that realm" to improve fluency and read- ability. 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 alterna- tive 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 in- tent, 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 "far- fetched" 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 ru- mored to halt aging in the Divine Realm." This hybrid "foreignization plus domestication" strat- egy balances cultural retention with reader com- prehension, 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 difficulty- adaptive Long CoT data is provided in Figure 10. The prompt employed for the quantitative evalua- tion 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 cor- responding 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.Based on this difficulty level, they allocate appropriate time and cognitive effort: •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 SentenceReference TranslationDifficultyReason “我只是想要听您的意见。”她 用一副恭敬的口吻说道。 “I just want to hear your opinion,” she said in a respectful tone. EasyA simple sentence with every- day vocabulary and direct gram- mar. No specialized knowledge re- quired. At the same time the waves are fan- ning 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 (*disper- sion*), requiring moderate struc- tural and lexical control. Die folgende Tabelle zeigt die Häu- figkeit und Schwere unerwünschter Wirkungen, die in >5% von 168 Pa- tienten mit Mesotheliom berichtet wurden... The table below provides the fre- quency and severity of undesir- able effects reported in>5% of 168 mesothelioma patients... HardLong and syntactically complex, includes multiple subordinate clauses, verb-final word order, anddomain-specificmedical 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 givensrc_langsentence intotgt_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 special- ized 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 con- tent, 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 differ- ences. •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 redun- dancy 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(562tokens): <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(476tokens): <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 chargé d '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(648tokens): <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. "chargé d'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> 153tokens 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> 182tokens Figure 15: Output examples of our TwT-7B. 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> 278tokens Prompt:Translate the following sentence from German into English. i) 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>(i) 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> 310tokens Figure 16: Output examples of our TwT-7B. 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> 556tokens Figure 17: Output examples of our TwT-7B. Method LawsNewsScienceSubtitles BLEUCOMETKIWITokensBLEUCOMETKIWITokensBLEUCOMETKIWITokensBLEUCOMETKIWITokens Large Language Models DeepSeek-V359.5889.1584.42-35.8686.9485.44-33.5488.2885.46-23.6682.4482.69- Gemini-2.0-Flash56.4388.9084.39-36.3986.3785.14-33.4787.6985.24-24.6481.9182.33- GPT-4o48.5088.2884.52-34.5886.2285.02-32.0987.3484.84-24.9181.9181.93- Large Reasoning Models DeepSeek-R1 59.3489.1584.6857733.3586.6785.4049831.8288.0885.4747820.5082.1082.84514 Gemini-2.0-Flash-Thinking55.7588.6184.2770233.9086.0284.79114931.9887.4785.06109223.8380.9081.22708 OpenAI-o3-mini42.7387.6784.4442832.9086.2985.3544331.3887.4285.2438522.9581.9082.40355 OpenAI-o148.1188.3285.1347833.9386.6185.4740832.3287.9285.5636722.9682.3182.60340 GPT-554.0689.3485.4378435.0586.6585.5274032.2188.0685.4760623.2382.3282.68519 QwQ-32B43.7787.3584.3966733.2485.7784.6358432.0587.1184.4256322.5881.5481.29584 MT-Specialized Models SFT-Parallel-7B56.6488.8284.28-28.8985.4383.91-29.2886.1983.68-26.9381.4180.26- ALMA-7B-R36.3485.4581.86-24.5383.9581.64-22.1684.3381.83-18.5680.1879.40- ALMA-13B-R40.3386.8983.10-26.8784.6582.43-24.3785.7282.59-19.5481.1180.00- TowerInstruct-7B-v0.250.5388.2582.96-30.7284.6182.45-27.7685.6082.99-22.4580.7580.01- TowerInstruct-13B-v0.152.2888.5783.09-32.2485.3683.09-29.4386.0083.17-23.7481.1980.74- CoT-FT-7B57.0588.7684.355130.1785.2383.314228.7286.0283.483928.1780.9279.2029 MT-R1-Zero-7B35.4986.7884.517231.6986.0184.546429.4586.8784.616122.4881.7381.7055 SSR-X-Zero-7B38.8886.4483.445628.1485.8783.985228.5487.1184.654923.1880.5081.9939 mExTrans-7B38.7787.1184.4459725.0586.0785.3455325.7487.3485.3154614.3181.4082.85476 Our Models TwT-Qwen2.5-7B-Instruct52.5688.7684.7331033.6386.7584.8831132.4187.5384.7729424.7482.6381.73247 TwT-Qwen2.5-14B-Instruct56.0389.0784.6532034.5786.6584.6328532.7587.5484.6627224.4482.7381.74241 Method LiteraryITKoranMedicalAverage BLEU COMET KIWI Tokens BLEU COMET KIWI Tokens BLEU COMET KIWI Tokens BLEU COMET KIWI Tokens BLEU COMET KIWI Tokens Large Language Models DeepSeek-V316.1477.2877.07-38.1683.9078.86-17.9474.9180.33-41.4384.1581.90-33.2983.3882.02- Gemini-2.0-Flash18.3777.2076.67-37.9383.2078.49-19.7074.9779.71-44.3984.4981.76-33.9183.0981.72- GPT-4o17.7577.4877.18-37.2383.5278.37-17.6575.0480.64-41.8984.2381.89-31.8383.0081.80- Large Reasoning Models DeepSeek-R111.2575.4475.1657436.6683.4978.6859317.0574.7680.6379040.6983.9482.3966731.3382.9581.91586 Gemini-2.0-Flash-Thinking18.0177.0676.6378137.2682.9478.4034519.5775.0379.7967743.3684.1481.7041532.9682.7781.48734 OpenAI-o3-mini17.4876.7776.6754637.1382.6378.2034315.5273.5480.1951139.8283.6081.7934629.9982.4881.78420 OpenAI-o116.5477.5077.8152136.4083.2079.7540316.7675.0382.0250640.4483.8883.2944130.9383.1082.70433 GPT-515.0176.7277.0885936.9283.5879.2249218.8375.4981.8075143.0084.2983.1353132.2983.3182.54660 QwQ-32B12.9975.9876.3186321.9382.9579.4358312.3673.5981.4796333.9583.2782.8273526.6182.2081.85693 MT-Specialized Models SFT-Parallel-7B15.7776.7375.37-40.6484.0879.35-21.0974.6678.39-43.7684.4082.47-32.8882.7280.96- ALMA-7B-R13.4675.1274.97-33.7581.1277.72-14.0971.9279.12-37.9783.0481.36-25.1180.6479.74- ALMA-13B-R14.2075.8975.81-34.2281.6377.78-14.7572.7679.63-40.3483.3881.49-26.8381.5080.35- TowerInstruct-7B-v0.215.5475.4974.93-38.2083.4778.67-10.9969.0270.13- 46.8184.5880.80-30.3881.4779.12- TowerInstruct-13B-v0.116.8176.0175.46-39.2683.8578.44-11.4169.2369.16-48.5584.9980.93-31.7281.9079.26- CoT-FT-7B15.3176.4574.965240.4183.7979.273520.5274.0478.164544.2184.3482.594633.0782.4480.6642 MT-R1-Zero-7B13.7576.9376.726934.5182.7179.595613.0572.8480.667127.2883.3182.697125.9682.1581.8865 SSR-X-Zero-7B13.6676.8276.145427.0180.0876.573613.9772.8679.754628.8682.1280.985025.2881.4780.9448 mExTrans-7B10.4376.1676.6261025.5478.7677.1945211.8073.8880.5660425.4081.6681.9356522.1381.5581.78551 Our Models TwT-Qwen2.5-7B-Instruct17.9378.3577.3028140.1284.0480.0522221.1375.4580.1926943.7484.3282.8126233.2883.4882.06274 TwT-Qwen2.5-14B-Instruct19.1078.4076.9335441.2684.1779.6823422.2075.7180.1333644.4584.4282.8928734.3583.5981.91291 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. Method ConversationEcommerceSocial BLEUCOMETKIWITokensBLEUCOMETKIWITokensBLEUCOMETKIWITokens Large Language Models DeepSeek-V336.7687.0381.48-32.2985.5381.19-32.2684.5981.44- Gemini-2.0-Flash38.0586.9081.37-32.4185.4980.87-32.9884.2881.04- GPT-4o38.0286.7981.43-32.8885.5781.04-32.7584.4781.08- Large Reasoning Models DeepSeek-R133.3486.4781.3753427.4185.0380.9155227.2283.9481.17554 Gemini-2.0-Flash-Thinking37.2186.6481.24120431.5085.2480.7482231.7983.8980.831081 OpenAI-o3-mini36.6686.3181.1529031.5185.1380.9736331.7984.0080.96372 OpenAI-o135.5087.1082.8132729.6585.6182.1339929.2184.5082.43405 GPT-535.2587.1482.8244829.0985.3782.0260929.0084.1082.31652 MT-Specialized Models SFT-Parallel-7B32.2084.1180.62-26.9482.9280.12-25.8881.6180.12- ALMA-7B-R29.4984.6779.96-25.1382.9879.50-25.7682.5179.77- ALMA-13B-R31.8685.4880.76-26.2483.7580.12-27.0083.0680.47- CoT-FT-7B31.5884.0980.633126.7182.9580.114524.9881.3779.8342 MT-R1-Zero-7B32.3885.8881.505326.9884.2581.256626.3983.2881.4065 SSR-X-Zero-7B30.6585.3281.133725.6483.9480.795025.4783.1981.2649 mExTrans-7B24.5384.7281.1546420.4683.5281.2456619.6082.2981.70555 Our Models TwT-Qwen2.5-7B-Instruct35.1086.3781.6923130.9284.8581.3527330.7984.0881.58269 TwT-Qwen2.5-14B-Instruct35.3586.3981.5824031.0485.1181.2130931.0483.9181.46298 Method CultureCommonSenseAverage BLEUCOMETKIWITokensBLEUCOMETKIWITokensBLEUCOMETKIWITokens Large Language Models DeepSeek-V340.2785.4683.21-32.5585.3679.98-34.8385.5981.46- Gemini-2.0-Flash39.1185.0282.94-31.6084.7179.38-34.8385.2881.12- GPT-4o38.6585.2483.14- 32.5485.2679.87-34.9785.4781.31- Large Reasoning Models DeepSeek-R135.88 85.4583.1456028.5084.5279.9960230.4785.0881.32561 Gemini-2.0-Flash-Thinking36.9484.6383.68122032.1584.5079.68233533.9284.9881.231332 OpenAI-o3-mini34.0584.3684.2559628.2384.2880.8543632.4584.8281.64411 OpenAI-o134.5785.0684.2554228.4384.9780.9539231.4785.4582.52413 GPT-535.7985.0484.1598427.0884.4380.7553031.2485.2282.41645 MT-Specialized Models SFT-Parallel-7B31.8182.8781.35-22.1981.9178.98-27.8182.6880.24- ALMA-7B-R33.2583.8182.83-23.2582.8080.00-27.3883.3580.41- ALMA-13B-R22.2381.1279.48-27.3582.4178.97-26.9483.1679.96- CoT-FT-7B30.4182.4681.105422.3881.9778.883327.2182.5780.1141 MT-R1-Zero-7B31.8683.9982.857924.5082.9479.535128.4284.0781.3163 SSR-X-Zero-7B28.7083.2581.116623.9382.8279.803426.8883.7080.8247 mExTrans-7B27.1284.3083.9263118.6181.9279.2347022.0683.3581.45537 Our Models TwT-Qwen2.5-7B-Instruct35.2484.5983.6435229.0884.3080.0721932.2384.8481.67269 TwT-Qwen2.5-14B-Instruct37.7084.7183.2333329.8184.3679.9825932.9984.9081.49288 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. Method En→ZhZh→EnDe→En BLEUCOMETKIWITokensBLEUCOMETKIWITokensBLEUCOMETKIWITokens Large Language Models Qwen2.5-7B-Instruct35.0585.4483.11-12.6369.8870.22-28.8279.3680.53- Gemma-2-9B-IT32.6684.4982.36-13.5772.4171.94-25.8477.0679.01- MT-Specialized Models ALMA-7B-R26.6984.1381.43-13.4675.1274.97-28.8779.6680.13- Tower-Plus-9B37.9186.7684.88-16.5677.8877.78-36.5682.2182.36- SFT-Parallel-7B38.5886.3483.39-15.7776.7375.37-35.7281.6381.11- MT-R1-Zero-7B32.0086.2384.346213.7576.9376.726925.5680.3981.7366 SSR-X-Zero-7B34.20 86.8684.585014.8377.4577.275426.6580.3981.5545 mExTrans-7B28.6386.4485.0253710.4376.1676.6261022.0679.2780.88544 Our Models TwT-Qwen2.5-7B-Instruct 38.2887.2784.4329817.9378.3577.3028135.5381.9381.90256 TwT-Gemma-2-9B-IT36.3486.6584.2322719.1778.1177.0724936.7382.1581.97218 Method En→X→EnAverage BLEUCOMETKIWITokensBLEUCOMETKIWITokensBLEUCOMETKIWITokens Large Language Models Qwen2.5-7B-Instruct5.9755.5651.30-20.4576.2274.30-20.5873.2971.89- Gemma-2-9B-IT14.0575.6373.16-31.5584.7182.88-23.5378.8677.87- MT-Specialized Models ALMA-7B-R2.7858.1373.13-12.0163.9458.17-16.7672.2073.57- Tower-Plus-9B6.0763.8669.46-27.9981.3879.68-25.0278.4278.83- SFT-Parallel-7B1.1952.1643.97-18.4775.5873.05-21.9574.4971.38- MT-R1-Zero-7B5.9158.7658.2735821.9177.4875.617419.8275.9675.34126 SSR-X-Zero-7B6.0258.9456.3430620.7378.0575.784820.4976.3475.10101 mExTrans-7B5.0961.3361.74104715.9576.8075.2173116.4376.0075.89694 Our Models TwT-Qwen2.5-7B-Instruct5.9959.5757.8348322.9477.7375.8432824.1376.9775.46329 TwT-Gemma-2-9B-IT11.1272.0280.8228032.7985.0283.3025727.2380.7981.48246 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.