Pith. sign in

REVIEW 4 major objections 3 minor 24 references

A 14B translation model with difficulty-adaptive reasoning can match or beat far larger reasoning models while using far fewer tokens.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 09:58 UTC pith:ZL3KWCLN

load-bearing objection Efficiency result is real; the "outperform in quality" claim is not supported by the paper's own metrics and human eval. the 4 major comments →

arxiv 2607.29287 v1 pith:ZL3KWCLN submitted 2026-07-31 cs.CL cs.AI

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

Yongshi Ye , Biao Fu , Chongxuan Huang , Yidong Chen , Xiaodong Shi This is my paper
classification cs.CL cs.AI
keywords multi-domain machine translationdifficulty-adaptive reasoningchain-of-thought distillationreinforcement learning for translationtranslation quality metricsresource-rational reasoningout-of-domain generalizationlow-resource translation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper's central claim is that translation quality and reasoning efficiency are not in tension: a model that learns to match its thinking depth to input difficulty gets both. TwT is trained in two stages — supervised fine-tuning on roughly seven thousand chain-of-thought traces that a teacher model generated and an annotator rewrote to the difficulty of each sentence, then reinforcement learning with a reward that combines BLEU and COMET quality with a repetition penalty. On a composite Quality score (the mean of BLEU, COMET, and CometKiwi) across eight in-domain domains, five out-of-domain domains, and 62 language directions, the 7B and 14B TwT models match or exceed much larger state-of-the-art reasoning models while emitting 32–60% fewer reasoning tokens. A sympathetic reading is that resource-rational reasoning — short checks for easy inputs, deeper deliberation for hard ones — is the mechanism that makes small models competitive.

Core claim

The claim is that reinforcement learning can bridge System 1 and System 2 translation behavior. The paper shows that pure RL collapses into repetitive templates and that uniform chain-of-thought over-thinks easy inputs, whereas the two-stage TwT recipe produces models that use roughly 200–300 tokens on easy inputs and 300–380 on hard ones while holding quality at or above the level of much larger reasoning models. TwT-14B's average Quality of 66.62 across the full evaluation suite is the headline number; the authors interpret it as evidence that aligning reasoning effort with input difficulty yields both efficiency and quality.

What carries the argument

The mechanism is a difficulty-adaptive chain-of-thought dataset plus a hybrid outcome reward. In the cold-start stage, a domain-specialized teacher generates long reasoning traces, an annotator grades each input as easy/medium/hard along four linguistic dimensions and rewrites the trace to that depth, and the backbone is fine-tuned on about 7K such examples. In the RL stage, GRPO optimizes a reward r = format reward + (BLEU + COMET) − λ·r_rep, where r_rep penalizes repeated 20-grams in the reasoning trace. The difficulty labels drive token allocation; the BLEU+COMET pair keeps lexical fidelity and semantic adequacy moving together; the repetition penalty prevents template collapse and over-t

Load-bearing premise

The load-bearing premise is that the composite Quality score — the mean of BLEU, COMET, and CometKiwi — is an unbiased measure of translation quality; since BLEU and COMET are also the reward terms TwT is trained on, the reported gap over DeepSeek-R1 could partly measure optimization of the evaluation metric rather than human-valued quality, and the paper's own 100-example human evaluation shows TwT-14B losing to DeepSeek-R1 (0.41 vs 0.47).

What would settle it

A larger, preregistered human evaluation on the same benchmarks (or an evaluation with a reference-based metric not used in training) that fails to place TwT-14B at parity with DeepSeek-R1 would refute the central outperformance claim. The paper already contains a directional clue: on 100 Zh↔En samples, professional translators preferred DeepSeek-R1 over TwT-14B 47% to 41%. Scaling that comparison across all domains would settle the question.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Small, efficient models can plausibly replace much larger reasoning models for practical multi-domain translation at lower inference cost.
  • The difficulty-adaptive behavior transfers out of domain and to 59 unseen languages, so the efficiency gain is not domain memorization.
  • Coupling lexical and semantic metrics in the reward appears to suppress reward hacking that occurs when a semantic metric is optimized alone.
  • The recipe generalizes across backbone architectures: Qwen2.5-7B, Qwen2.5-14B, Llama-3.1-8B, and Gemma-2-9B all improve over their base models, indicating the mechanism, not scale, carries the gain.
  • The small data budget (about 27K training examples) suggests low-resource translation settings can bootstrap adaptive reasoning without massive parallel corpora.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A reader should not treat the automatic-metric lead as proof of human-preferred quality: the paper's own human evaluation on 100 examples shows TwT-14B losing to DeepSeek-R1 (0.41 vs 0.47), which is a direct warning that the composite Quality score may partially reflect the BLEU and COMET terms used in training.
  • The difficulty labels come from GPT-4o and are validated on 100 samples (Appendix F.7); an open, larger validation set would make the adaptive-reasoning claim more reproducible and would test how sensitive TwT is to label noise.
  • A natural next step, which the paper flags as future work, is difficulty-aware reward shaping: reward concise reasoning on easy inputs and deeper reasoning on hard ones. That would likely widen the token-saving gap and make the resource-rational behavior even more explicit.
  • The same difficulty-adaptive frame could be tested on other generation tasks with variable input complexity, such as summarization, code repair, or question answering, where uniform overthinking is also a cost.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. The paper proposes TwT, a resource-rational framework for multi-domain machine translation that adapts chain-of-thought reasoning depth to input difficulty. Training uses cold-start SFT on difficulty-adaptive CoT traces (distilled from DeepSeek-R1, rewritten by GPT-4o) followed by GRPO RL with a reward combining BLEU, COMET, a format reward, and a repetition penalty. Evaluation covers eight in-domain domains, five OOD domains, and 62 language directions, with the headline result that TwT-7B/14B outperform much larger reasoning models on a composite Quality score while using 32–60% fewer tokens.

Significance. The demonstrated ability to reduce reasoning tokens while maintaining or improving automatic MT scores is practically relevant, and the three-backbone ablation, full metric breakdowns, and detailed training details are strengths. The paper is transparent about data sources, hyperparameters, and limitations. However, because two of the three Quality components are exactly the RL reward terms, and the paper's own human evaluation shows TwT-14B losing to DeepSeek-R1 and DeepSeek-V3, the central claim of 'outperform ... in translation quality' is not yet established. The contribution may still be valuable after reframing the claim around efficiency and competitive automatic metrics.

major comments (4)
  1. [§5.1, §4.2, Table 18] The headline Quality metric is the mean of BLEU, COMET, and CometKiwi, but the RL quality reward is r_q = B + C (BLEU + COMET). Two of the three metric components are therefore the training objective. Table 18 shows that TwT-14B's in-domain edge over DeepSeek-R1 is BLEU 34.35 vs 31.33, COMET 83.59 vs 82.95, CometKiwi 81.91 vs 81.91 (tie). The 1.22-point Quality advantage comes entirely from the rewarded components. Consequently the abstract's claim of 'outperform much larger SOTA reasoning models in translation quality' is not supported as stated; the comparison partly measures reward optimization. Please provide the primary result on metrics not in the reward (e.g., CometKiwi alone, MQM, human) or clearly qualify the claim.
  2. [Table 16 / §G.1] The paper's own human evaluation of 100 Zh↔En examples shows TwT-14B wins 41% vs DeepSeek-R1 47% (tie 12%) and loses to DeepSeek-V3 (27.5% vs 40.5%). The text says this is 'expected given its much smaller size' and claims 'strong degree of parity,' but the Abstract and §5.2 state TwT-14B 'outperforms' DeepSeek-R1 and GPT-5. These statements are in direct tension. Please reconcile the automatic-metric headline with the human evaluation, e.g., by reporting human evaluation as the primary quality criterion or by explicitly limiting the claim to automatic composite metrics.
  3. [Tables 3–5, 8] Differences are reported without confidence intervals or significance tests. For example, TwT-14B vs GPT-5 (66.62 vs 66.05) and vs DeepSeek-R1 (66.62 vs 65.40) are small relative to the scale of the metric, and Table 18 shows component-wise variation. Given the composite and reward overlap, the paper should report bootstrap CIs or paired significance tests for the core comparisons. This is essential to establish the claimed superiority.
  4. [§4.1, Appendix F.7] The difficulty-adaptive training data are generated using GPT-4o difficulty labels, but the validation of these labels is based on 100 randomly sampled instances (F.7). This is a very small sample for estimating agreement with human experts, and the paper does not report per-dimension agreement or the label distribution across domains. Since the entire two-stage pipeline hinges on the reliability of these labels, a larger validation or an analysis of label noise's effect on downstream performance is needed.
minor comments (3)
  1. [Table 8 / Figure 11] In §6.3 you state the difficulty grouping is 'estimated by DeepSeek-V3 (prompt in Figure 11)', but Figure 11 shows the GPT-4o difficulty prompt used in §4.1. Please clarify which model produced the difficulty labels for Table 8.
  2. [Table 16 / §G.1] Report inter-annotator agreement for the human evaluation; with three translators and no agreement metric, the 0.41 vs 0.47 result is hard to interpret.
  3. [Limitations] The Limitations section does not discuss the overlap between the evaluation metric and the reward; adding a sentence would help calibrate readers.

Circularity Check

1 steps flagged

Headline quality claim is measured partly by the RL reward metrics; the paper's own human eval reverses it.

specific steps
  1. fitted input called prediction [Abstract; §4.2 RL Training; §5.1 Metrics; Appendix I Table 18]
    "TwT-7B and TwT-14B outperform much larger SOTA reasoning models in translation quality... We report Quality, defined as the average of BLEU, COMET (Rei et al., 2020), and CometKiwi... we adopt a hybrid quality reward that combines BLEU (B) with COMET (C): rq = ( B(ŷ,y) + C(x,ŷ,y) if rf = 1 ...)"

    Quality is introduced as the yardstick for the abstract's 'translation quality' claim: 'We report Quality, defined as the average of BLEU, COMET, and CometKiwi' (§5.1). But the RL stage optimizes exactly two of these three components: 'we adopt a hybrid quality reward that combines BLEU (B) with COMET (C): rq = B(ŷ,y)+C(x,ŷ,y)' (§4.2). Up to normalization, Quality = (r_q + CometKiwi)/3, so the headline comparison is partly a measure of the training objective. Table 18 shows the entire in-domain margin over DeepSeek-R1 is BLEU (34.35 vs 31.33) plus COMET (83.59 vs 82.95), with CometKiwi tied (81.91 vs 81.91). The paper's own human evaluation (Table 16) shows TwT-14B loses to DeepSeek-R1 (0.41 vs 0.47) and DeepSeek-V3 (0.275 vs 0.405), so the 'outperform' claim is not independently confirmed

full rationale

TwT's two-stage training is internally coherent: cold-start SFT on distilled difficulty-adaptive traces and GRPO with hybrid reward are externally anchored to DeepSeek-R1, GPT-4o, and standard MT benchmarks. The ablation study and multilingual generalization are evaluated on held-out test sets, and the token-efficiency claim is supported by direct token counts independent of the reward. The central circularity is limited to the headline quality claim: Quality = mean(BLEU, COMET, CometKiwi) shares BLEU and COMET with the RL reward r_q = B + C, so comparing against models not trained on this reward is partly measuring the training objective. This is confirmed by Table 18 (the DeepSeek-R1 edge is almost entirely BLEU) and by Table 16, where TwT-14B loses to DeepSeek-R1 and DeepSeek-V3 in human evaluation. No load-bearing self-citation or imported uniqueness theorem is present. The 100-sample GPT-4o difficulty validation is a reliability concern, not circularity. Score 6 reflects partial, not total, circularity: CometKiwi and held-out test sets provide some independent signal, and the efficiency claim is not circular.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

No invented physical/conceptual entities: 'System 1/System 2' and 'Rich Points' are borrowed from prior literature (Hvelplund 2011; Agar 1994). The ledger's real content is proxy choices: token length as cognitive effort, GPT-4o difficulty as ground truth, and BLEU+COMET as both the training reward and part of the evaluation metric. The last is the load-bearing one for the headline claim.

free parameters (4)
  • λ (repetition penalty weight) = not stated
    r = r_f + r_q − λ·r_rep (§4.2); the weight that balances translation quality against the token-efficiency claim (32-60% reduction) is never given in the text, so the efficiency results cannot be re-derived.
  • n (n-gram window for repetition penalty) = 20
    Set by hand in §4.2; controls what counts as a repetitive loop and thus the magnitude of r_rep.
  • β (KL coefficient) = 1e-3
    Chosen after a sweep (Table 13) showing β=0.001/0.005 best; the reported value sits near the good end of that sweep (in-domain Quality 64.77 vs 64.85 at β=0.005).
  • GPT-4o difficulty thresholds (easy/medium/hard boundaries) = uncalibrated rubric
    The four-dimension rubric (Figure 11) has no anchor definitions; boundaries between easy/medium/hard are validated only on 100 samples (Appendix F.7; corr 0.86).
axioms (6)
  • ad hoc to paper Reasoning depth is equivalent to CoT token length
    §4.1: 'We define reasoning depth as the length of the generated reasoning trace (i.e., number of tokens)'. Equates cognitive effort with generated tokens; no independent evidence.
  • domain assumption GPT-4o difficulty labels are a valid proxy for human translation difficulty
    The whole adaptive-rewriting pipeline (cold start) conditions trace depth on these labels; validated only on 100 instances (Appendix F.7, corr 0.86, 90% cross-model agreement).
  • domain assumption BLEU+COMET hybrid reward does not induce reward hacking
    §3.1 shows CometKiwi-only reward hacks (BLEU collapses); §4.2 asserts BLEU+COMET stabilizes. Training curves (Figure 8) support stability, but the human eval (Table 16) suggests the resulting quality gain over DeepSeek-R1 is not detectable by humans.
  • domain assumption The composite Quality score is an unbiased measure of translation quality
    §5.1 defines Quality = mean(BLEU, COMET, CometKiwi); BLEU and COMET are exactly TwT's RL rewards (§4.2), so the metric is partially the training objective for TwT but not for closed baselines.
  • domain assumption DeepSeek-R1 reasoning traces are high-quality translation demonstrations
    Cold-start SFT data is distilled from DeepSeek-R1 with domain-aware prompts (§4.1); no independent quality audit of traces beyond the GPT-4o rewriting (which checks style/difficulty, not correctness).
  • ad hoc to paper The four difficulty dimensions (complexity, vocabulary rarity, grammatical divergence, contextual nuance) span translation difficulty
    Defined in the GPT-4o evaluation prompt (Figure 11); no prior literature cited justifying completeness.

pith-pipeline@v1.3.0-daily-deepseek · 38245 in / 20474 out tokens · 194659 ms · 2026-08-03T09:58:51.618881+00:00 · methodology

0 comments
read the original 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.

Figures

Figures reproduced from arXiv: 2607.29287 by Biao Fu, Chongxuan Huang, Xiaodong Shi, Yidong Chen, Yongshi Ye.

Figure 1
Figure 1. 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 align… view at source ↗
Figure 2
Figure 2. Figure 2: Training dynamics under pure RL using different quality rewards. While translation quality improves [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of TwT training pipeline. TwT is first fine-tuned on difficulty-adaptive Long CoT traces distilled from DeepSeek-R1 and rewritten by GPT-4o for cognitive alignment. RL is then applied with a hybrid reward. locate deep reasoning selectively—only where it leads to measurable quality improvements. By mit￾igating indiscriminate overthinking (Section 3.2), we improve reasoning efficiency and align the … view at source ↗
Figure 4
Figure 4. Figure 4: CoT trace similarity comparison. Pure RL vs [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Template for pure RL in MT task. • The German-English multi-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 tas… view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of our curated difficulty-adaptive Long CoT data. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Average reward and format error rate over RL [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. 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 [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Inference time comparison. ferent models. As shown in [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Full prompt used for generating difficulty-adaptive Long CoT data with GPT-4o. [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Full prompt used for evaluating translation difficulty with GPT-4o. [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The prompt used for the quantitative evaluation of reasoning efficiency. [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Examples of our curated difficulty-adaptive Long CoT data (Easy and Medium). [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Examples of our curated difficulty-adaptive Long CoT data (Hard). [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Output examples of our TwT-7B [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Output examples of our TwT-7B [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Output examples of our TwT-7B [PITH_FULL_IMAGE:figures/full_fig_p029_17.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 3 linked inside Pith

  1. [1]

    Sentence complexity: Determine if the sentence is a simple sentence, a compound sentence, or includes subordinate clauses and other complex structures

  2. [2]

    Jie He, Tao Wang, Deyi Xiong, and Qun Liu

    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. InFindings of the Association for Computational Linguistics: EMNLP 2020, pages 3662–3672, Online. Associat...

    Pith/arXiv arXiv 2020

  3. [3]

    Translate the following {src_lang} text into {tgt_lang}

    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...

    Pith/arXiv arXiv 2024

  4. [4]

    easy", "medium

    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 "r...

  5. [5]

    cited Vietnamese data

    "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. [6]

    V ocabulary difficulty: Assess whether the sentence contains commonly used words or special- ized terms or slang

  7. [7]

    Grammar differences: Analyze if the sentence’s grammatical structure is similar to or differs significantly from{tgt_lang}

  8. [8]

    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。

    引号方面，层层嵌套时需注意区分人物直述与转述，建议采用中文标准引号和冒号。译文顺序可适当调整以更贴近中文表达习惯。 完成初稿后，逐句核对语气严谨、表达完整，查漏补缺，确保术语统一和风格正式。再次复核关键历史用词、机构名称准确无误，并确保译文逻辑通顺。</think> <answer>1956年6月15日，越南民主共和国外交部副外长雍文谦在会见中国驻越南使馆临时代办李志民时郑重表示：“根据越南方面的资料，西沙群岛和南沙群岛自古以 来就是中国的领土。”当时在场的越南外交部亚洲司代司长黎禄还专门引用了越南方面的资料指出：“从历史上看，这些群岛在宋代时已经属于中国。”</answer> Figure 14: Examples of our curated difficulty-adaptive Long CoT ...

    arXiv 2079

  9. [9]

    外门弟子”为宗门中层级较低的弟子，常见翻译为“outer sect disciples

    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”。

  10. [10]

    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

    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。

  11. [11]

    现在他要找的就是……

    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...”。

  12. [12]

    thought" and

    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 r...

  13. [13]

    Now, he needed to find the place where the outer sect disciples cultivated

    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:...

    1956

  14. [14]

    chargé d'affaires ad interim

    "chargé d'affaires ad interim" 正确翻译是“临时代办”，注意头衔的准确。

  15. [15]

    solemnly stated

    "solemnly stated" 要译为“郑重声明”，符合外交场合的正式用语。

  16. [16]

    Vietnamese data

    "Vietnamese data" 翻译为“越南方面的材料”比较合适，避免直译成“数据”。

  17. [17]

    Acting Director

    "Acting Director" 是“代理司长”。

  18. [18]

    Song Dynasty

    "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...

    1956

  19. [19]

    chargé d'affaires ad interim

    “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”。

  20. [20]

    During a meeting ... on 15 June 1956

    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”。

    1956

  21. [21]

    Vietnamese data

    “Vietnamese data”一般译为“越南方面资料”；

  22. [22]

    acting director

    “acting director”常见译法为“代司长”，“Le Loc”应对照权威译名为“黎禄”。

  23. [23]

    judging from history

    “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。

  24. [2025]

    Preprint, arXiv:2502.11544

    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. Ge...

    Pith/arXiv arXiv 2024