Pith. sign in

REVIEW 3 major objections 5 minor 66 references

TACTIC: Translation Agents with Cognitive-Theoretic Interactive Collaboration

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a six-agent cognitive workflow, which drafts, merges, critiques, and iteratively reworks translations, reliably beats one-shot LLM translation across model families, language pairs, and two benchmarks.

desk verdict Solid multi-agent MT recipe with reproducible same-model gains, but the SOTA claim is undermined by missing compute-matched baselines and a questionable reverse-direction evaluation. read the letter →

arxiv 2506.08403 v2 pith:TNQ5TNVO submitted 2025-06-10 cs.CL cs.AI

Weiya Li , Junjie Chen , Bei Li , Boyang Liu , Zichen Wen , Nuanqiao Shan , Xiaoqian Liu , Anping Liu
show 3 more authors
Huajie Liu Hu Song Linfeng Zhang
This is my paper
classification cs.CLcs.AI
keywords machinetranslationmulti-agentsystemslargelanguagemodelscognitivestudiesiterativerefinementevaluationworkfloworchestration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that machine translation quality can be pushed past what a single LLM call produces by simulating the full cognitive workflow of a human translator: generating several stylistically different drafts, merging them, critiquing the result on faithfulness, expressiveness, and elegance, and only then judging whether to accept it. If the result fails the bar, the system escalates to a complex workflow where two extra agents research terminology and reconstruct the discourse context, then the cycle repeats. Across five language pairs on FLORES-200 and WMT24, the authors report consistent gains over zero-shot and few-shot baselines, with DeepSeek-V3 as the engine outperforming GPT-4.1 by +0.6 XCOMET and +1.18 COMETKIWI-23 on average. A sympathetic reader would take the paper's message as: the bottleneck in LLM translation is not model capability alone but how the model is orchestrated, and a theory-grounded agent architecture can unlock that headroom.

What carries the argument

The load-bearing mechanism is the threshold-gated escalation loop: a base workflow that drafts in three cognitive styles, refines, evaluates on faithfulness, expressiveness, and elegance, and scores; when the score falls below a preset threshold, the workflow shifts to a complex loop in which a ResearchAgent mines keywords and phrases and a ContextAgent reconstructs surrounding discourse before a new draft-refine-evaluate cycle begins. The coupling of a fast path for routine sentences and a slow reflective path for difficult ones is what the paper claims converts raw LLM capability into better translation.

What would settle it

Take the strongest setup (DeepSeek-V3) and run two controls on the same language-pair and benchmark conditions: best-of-N sampling with N equal to TACTIC's average LLM-call count, and a plain self-refinement loop with the same call budget. If either control matches or exceeds TACTIC's XCOMET and COMETKIWI-23 scores, then the six-agent structure is not the cause of the gain; if both fall short by the same margin as the published baselines, the architecture is doing the work.

Watch

Extended reading notes

Core claim

TACTIC organizes six specialized agents — Draft, Refinement, Evaluation, Score, Context, and Research — into a two-layer workflow. The DraftAgent produces literal, sense-for-sense, and free translations, mirroring the divergent thinking phase of human translation; the RefinementAgent merges these into one coherent output; the EvaluationAgent assesses the result along the classical dimensions of faithfulness, expressiveness, and elegance; and the ScoreAgent converts the critique into a threshold decision. When the score falls below a quality threshold, the complex workflow activates the ResearchAgent and ContextAgent to supply terminology and contextual framing, and the drafting, refining, and evaluating cycle repeats until the threshold is met. The paper's central claim is that this architecture, grounded in Cognitive Translation Studies, consistently improves scores over direct prompting: it reports state-of-the-art results on FLORES-200 and WMT24, including DeepSeek-V3-based TACTIC surpassing GPT-4.1 zero-shot and DeepSeek-R1 baselines on both XCOMET and COMETKIWI-23.

Load-bearing premise

The reported gains are attributed to the six-agent cognitive workflow, yet the experiments do not compare against a control that spends the same amount of inference compute — every baseline makes one or a few LLM calls while TACTIC makes at least four and often many more — so the load-bearing premise is that the architecture, not the extra budget, produces the improvement.

Editorial extensions

If this is right

  • The same base model scores higher under TACTIC than under direct prompting across every tested model family, from Qwen2.5-7B to DeepSeek-V3, in both zero-shot and few-shot setups.
  • Because the workflow is model-agnostic, the reported gains transfer across language pairs from German to Chinese, across both translation directions, and across both benchmarks.
  • The ablation shows each added component — iterative evaluation, drafting-then-refining, and keyword/context mining — contributes a measurable increment, so the benefit is cumulative rather than coming from one agent alone.
  • The paper's analysis concludes that structured collaboration can outweigh intrinsic reasoning strength, since DeepSeek-V3 under TACTIC beats even DeepSeek-R1 used with the framework.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The framework opens a natural control experiment the paper does not run: matching TACTIC's per-sentence call budget with best-of-N sampling or a plain self-refinement loop to separate architectural gains from compute gains.
  • Because the ScoreAgent is an LLM whose threshold is not calibrated against human judgments, the escalation from base to complex workflow is a soft, model-dependent dial; tuning it could trade latency against quality in ways the paper leaves unexplored.
  • The metrics that carry the headline results are learned model-based scores, and the paper itself notes they disagree with lexical measures; a human evaluation on the same outputs would reveal which quality dimension the workflow actually improves.
  • If orchestration substitutes for model scale, the practical recipe implied by the results is small open-weight models wrapped in a similar loop, trading compute for cheaper deployment — a direction the paper does not test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes TACTIC, a multi-agent translation framework whose six agents (DraftAgent, RefinementAgent, EvaluationAgent, ScoreAgent, ContextAgent, ResearchAgent) are motivated by Cognitive Translation Studies. A base workflow generates multiple drafts, refines them, evaluates them on faithfulness/expressiveness/elegance, and scores the result; if the score falls below a threshold, a complex workflow adds research and context agents and iterates until the threshold or a stopping condition is met. The authors evaluate on FLORES-200 and WMT24 with XCOMET and COMETKIWI-23, reporting consistent gains over zero-shot and few-shot baselines and claiming state-of-the-art results when DeepSeek-V3 is used as the backend, including a +0.6 XCOMET improvement over GPT-4.1.

Significance. If the reported gains are real and attributable to the architecture, TACTIC would be a useful contribution to LLM-based translation: it provides a concrete operationalization of cognitive translation concepts, releases code, specifies its workflow in Algorithm 1, and uses external metrics (XCOMET, COMETKIWI-23) rather than the internal ScoreAgent judgments for the reported numbers, which reduces circularity concerns. The same-model comparisons in Table 2 (e.g., DeepSeek-V3 few-shot 94.90 vs. TACTIC 96.19 on FLORES en-xx) suggest a genuine improvement over a one-call baseline. However, the central attribution of these gains to the six-agent design is not yet supported because the experiments do not control for the substantially larger inference budget consumed by TACTIC.

major comments (3)
  1. [§3.2, Table 2, Algorithm 1] The reported gains are confounded with inference compute. The baseline systems make one LLM call per sentence, whereas TACTIC's base workflow invokes DraftAgent, RefinementAgent, EvaluationAgent, and ScoreAgent, and the complex workflow in Algorithm 1 (lines 10–17) repeats research, context, drafting, refinement, evaluation, and scoring calls until a threshold is met. No compute-matched baseline (e.g., best-of-N sampling from the same base model, or an equal-budget single-agent self-refinement loop) is included, so the improvement cannot be attributed to the cognitive architecture rather than to the much larger number of samples and refinement steps. Section 6 acknowledges higher latency but does not control for compute budget in the experiments.
  2. [§3.3, Table 3] The ablation study is also confounded with the number of LLM calls. Each row adds modules that add calls: '+ Iterative Evaluation' adds evaluation and scoring calls, 'Drafting-then-Refining' adds draft and refinement calls, and '++ Keyword, Phrase and Context Mining' adds ResearchAgent and ContextAgent calls. The monotonic improvements could therefore reflect additional inference compute rather than the specific cognitive roles. An ablation that holds the total token/call budget fixed, for example by reducing the number of drafts or the number of evaluation rounds in the control conditions, is needed to isolate the contribution of each agent.
  3. [§3.2 and Abstract] The headline state-of-the-art comparisons against GPT-4.1 and DeepSeek-R1 are not apples-to-apples: TACTIC is run with few-shot prompts and multiple agent calls, while those baselines are zero-shot single-call systems. Furthermore, no error bars or significance tests are reported anywhere in Table 2, despite Appendix B acknowledging randomness in LLM outputs. Differences around +0.6 XCOMET may be within run-to-run variation. The authors should report multiple runs with variance or paired significance tests, and compare against baselines under identical prompt settings and call budgets, before claiming consistent state-of-the-art performance.
minor comments (5)
  1. [Figure 3] The legend contains typos: 'Fist' should be 'First' and ''Last'' should be 'Last'.
  2. [§3.1] The text says the authors focus on three primary metrics, including MetricX-24-XXL, but Table 2 reports only XCOMET and COMETKIWI-23; MetricX-24 results appear only in Table 4. Please clarify which tables report which metrics.
  3. [Algorithm 1 vs. §B] Algorithm 1 shows the loop terminating only when s' ≥ τ, but Section B introduces maximum iteration threshold κ and maximum execution time threshold δ as stopping conditions. The algorithm should include these conditions so that the pseudocode matches the implementation.
  4. [Table 2] The COMETKIWI-23 value of 75.01 for DeepSeek-R1 on WMT24 xx→en is strikingly lower than neighboring entries (e.g., 80.45 for WMT24 en→xx and 88.36 for FLORES xx→en). Please verify this number or explain the anomaly.
  5. [§D, Table 6] Table 6 refers to 'Qwen3-32B model in non-inference mode'; 'non-inference mode' is undefined and the name Qwen3 is not used elsewhere in the paper. Please define the terminology and reconcile it with the model list in Section 3.1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are evaluated with external metrics, not with the framework's internal selection scores.

full rationale

This paper presents an empirical multi-agent translation system rather than a formal derivation, and none of the load-bearing claims reduce to the paper's own inputs by construction. The central result is measured on FLORES-200 and WMT24 using external reference-based and reference-free metrics (XCOMET, COMETKIWI-23, MetricX-24), while the TACTIC workflow's internal ScoreAgent threshold and iterative loop select outputs using the LLM's own self-assessment. Because the reported numbers come from metrics external to the agent configuration, the outcome is not defined in terms of the internal scores and there is no fitted parameter renamed as a prediction. The cognitive-translation framing is motivational and does not invoke an author-supplied uniqueness theorem or an ansatz smuggled in via citation. The one reference listing an author of this paper (Qiang Wang, Bei Li, et al., on deep transformer models) is a standard background citation and is not load-bearing. The absence of compute-matched baselines is a genuine experimental confound for causal attribution of the gains to the six-agent design, but it is a validity concern rather than a circularity: the comparison still uses external, independently defined quality measures. Therefore no circular step is exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical or metaphysical entities are introduced. The six agents are software roles, not independent entities with falsifiable handles. The key unstated assumptions are compute parity and the reliability of self-scoring, both load-bearing for the claimed gains.

free parameters (5)
  • quality threshold tau = not reported
    The ScoreAgent acceptance threshold in Algorithm 1 controls how many refinement iterations run; its value is never specified.
  • max iteration threshold kappa = not reported
    Stopping condition for the complex workflow in Appendix B; number of iterations is not reported, so the compute budget is unknown.
  • max execution time threshold delta = not reported
    Time-based stopping condition in Appendix B; not specified.
  • few-shot examples = 5 examples from Xu et al. (2024)
    Used in DraftAgent and RefinementAgent; chosen from prior work, not tuned here.
  • temperature = 0.6 (0.3 for DeepSeek-V3)
    Sampling temperature setting; not fit to data.
assumptions (5)
  • domain assumption CTS categories (cognitive strategies, cognitive processing, contextual cognition) provide a useful decomposition for MT agent design.
    Invoked in Section 1 and Table 1 to motivate the six agents; no evidence that this decomposition is optimal or that CTS is predictive of MT performance.
  • domain assumption LLM prompts can faithfully perform each agent role (drafting, refinement, evaluation, scoring, context expansion, research).
    Section 2.1 assumes the backend LLM follows role instructions and produces valid JSON; only anecdotal case studies support this.
  • domain assumption ScoreAgent's self-assessed quality correlates with true quality measured by XCOMET/COMETKIWI-23.
    Algorithm 1 selects outputs based on internal scores; the paper does not validate this correlation, yet iterative selection depends on it.
  • domain assumption Automatic metrics approximate human translation quality well enough to support SOTA claims.
    Limitations section admits reliance on automatic metrics; model-based metrics correlate with human judgments per prior work, but the paper provides no human evaluation.
  • domain assumption WMT24 provides only en-xx test sets for the five language pairs, so xx-en must be constructed by back-translating sources.
    Section 3.1 states this; if official WMT24 xx-en test sets exist, the reverse-direction results are not directly comparable to the benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TACTIC: Translation Agents with Cognitive-Theoretic Interactive Collaboration." pith.science (2026). https://pith.science/paper/TNQ5TNVO

@misc{pith2026250608403,
  author       = {Pith},
  title        = {Pith review of: TACTIC: Translation Agents with Cognitive-Theoretic Interactive Collaboration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TNQ5TNVO}},
  note         = {Machine review of arXiv:2506.08403}
}
read the original abstract

Machine translation has long been a central task in natural language processing. With the rapid advancement of large language models (LLMs), there has been remarkable progress in translation quality. However, fully realizing the translation potential of LLMs remains an open challenge. Recent studies have explored multi-agent systems to decompose complex translation tasks into collaborative subtasks, showing initial promise in enhancing translation quality through agent cooperation and specialization. Nevertheless, existing multi-agent translation frameworks largely neglect foundational insights from cognitive translation studies. These insights emphasize how human translators employ different cognitive strategies, such as balancing literal and free translation, refining expressions based on context, and iteratively evaluating outputs. To address this limitation, we propose a cognitively informed multi-agent framework called TACTIC, which stands for T ranslation A gents with Cognitive- T heoretic Interactive Collaboration. The framework comprises six functionally distinct agents that mirror key cognitive processes observed in human translation behavior. These include agents for drafting, refinement, evaluation, scoring, context reasoning, and external knowledge gathering. By simulating an interactive and theory-grounded translation workflow, TACTIC effectively leverages the full capacity of LLMs for high-quality translation. Experimental results on diverse language pairs from the FLORES-200 and WMT24 benchmarks show that our method consistently achieves state-of-the-art performance. Using DeepSeek-V3 as the base model, TACTIC surpasses GPT-4.1 by an average of +0.6 XCOMET and +1.18 COMETKIWI-23. Compared to DeepSeek-R1, it further improves by +0.84 XCOMET and +2.99 COMETKIWI-23. Code is available at https://github.com/weiyali126/TACTIC.

Figures

Figures reproduced from arXiv: 2506.08403 by the authors.

Figure 1
Figure 1. Overall agent collaboration workflow in the TACTIC framework. The figure depicts a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Case study demonstrating translation refinement through iterative agent collaboration. The [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Distributional Impact of Iterative Refinement on the WMT24 Test Set. “Fist” and “’Last” [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Case study visualizing different translation strategies and evaluation dimensions. The [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 41 canonical work pages

  1. [1]

    Tower: An open multilingual large language model for translation-related tasks

    Duarte M Alves, José Pombal, Nuno M Guerreiro, Pedro H Martins, João Alves, Amin Farajian, Ben Peters, Ricardo Rei, Patrick Fernandes, Sweta Agrawal, et al. Tower: An open multilingual large language model for translation-related tasks. arXiv preprint arXiv:2402.17733, 2024

    arXiv 2024

  2. [2]

    Neural machine translation by jointly learning to align and translate

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In Proc. of ICLR, 2015

    work page 2015

  3. [3]

    Translating step-by-step: Decomposing the translation process for improved translation quality of long-form texts, 2024

    Eleftheria Briakou, Jiaming Luo, Colin Cherry, and Markus Freitag. Translating step-by-step: Decomposing the translation process for improved translation quality of long-form texts, 2024. URL https://arxiv.org/abs/2409.06790

    arXiv 2024

  4. [4]

    Crat: A multi-agent framework for causality-enhanced reflective and retrieval-augmented translation with large language models, 2024

    Meiqi Chen, Fandong Meng, Yingxue Zhang, Yan Zhang, and Jie Zhou. Crat: A multi-agent framework for causality-enhanced reflective and retrieval-augmented translation with large language models, 2024. URL https://arxiv.org/abs/2410.21067

    arXiv 2024

  5. [5]

    Cognitive processes in translation and interpreting

    Giuseppina Cortese. Cognitive processes in translation and interpreting. joseph h. danks, gregory m. shreve, stephen b. fountain, and michael k. mcbeath (eds.). london: Sage, 1997. pp. 294. Applied Psycholinguistics, 20(2):318–327, 1999

    work page 1997

  6. [6]

    No language left behind: Scaling human-centered machine translation

    Marta R Costa-Jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672, 2022

    arXiv 2022

  7. [7]

    Wmt24++: Expanding the language coverage of wmt24 to 55 languages & dialects

    Daniel Deutsch, Eleftheria Briakou, Isaac Caswell, Mara Finkelstein, Rebecca Galor, Juraj Juraska, Geza Kovacs, Alison Lui, Ricardo Rei, Jason Riesa, et al. Wmt24++: Expanding the language coverage of wmt24 to 55 languages & dialects. arXiv preprint arXiv:2502.12404, 2025

    arXiv 2025

  8. [8]

    Tear: Improving llm-based machine translation with systematic self-refinement, 2024

    Zhaopeng Feng, Yan Zhang, Hao Li, Bei Wu, Jiayu Liao, Wenqiang Liu, Jun Lang, Yang Feng, Jian Wu, and Zuozhu Liu. Tear: Improving llm-based machine translation with systematic self-refinement, 2024. URL https://arxiv.org/abs/2402.16379

    arXiv 2024

Show all 66 references
  1. [9]

    Mt-r1-zero: Advancing llm-based machine translation via r1-zero-like reinforcement learning

    Zhaopeng Feng, Shaosheng Cao, Jiahan Ren, Jiayuan Su, Ruizhe Chen, Yan Zhang, Zhe Xu, Yao Hu, Jian Wu, and Zuozhu Liu. Mt-r1-zero: Advancing llm-based machine translation via r1-zero-like reinforcement learning. arXiv preprint arXiv:2504.10160, 2025

    2025 arXiv

  2. [10]

    M-mad: Multidimensional multi-agent debate for advanced machine translation evaluation, 2025

    Zhaopeng Feng, Jiayuan Su, Jiamei Zheng, Jiahan Ren, Yan Zhang, Jian Wu, Hongwei Wang, and Zuozhu Liu. M-mad: Multidimensional multi-agent debate for advanced machine translation evaluation, 2025. URL https://arxiv.org/abs/2412.20127

    2025 arXiv

  3. [11]

    Basic concepts and models for interpreter and translator training

    Daniel Gile. Basic concepts and models for interpreter and translator training. John Benjamins Publishing Company, 2009. 10

    2009

  4. [12]

    Building a translation competence model

    Pacte Group et al. Building a translation competence model. In Triangulating translation: Perspectives in process oriented research, pages 43–66. John Benjamins Publishing Company, 2008

    2008

  5. [13]

    xcomet: Transparent machine translation evaluation through fine-grained error detection

    Nuno M Guerreiro, Ricardo Rei, Daan van Stigt, Luisa Coheur, Pierre Colombo, and An- dré FT Martins. xcomet: Transparent machine translation evaluation through fine-grained error detection. Transactions of the Association for Computational Linguistics, 12:979–995, 2024

    2024

  6. [14]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  7. [15]

    Redefining machine translation on social network services with large language models

    Hongcheng Guo, Fei Zhao, Shaosheng Cao, Xinze Lyu, Ziyan Liu, Yue Wang, Boyang Wang, Zhoujun Li, Chonggang Lu, Zhe Xu, et al. Redefining machine translation on social network services with large language models. arXiv preprint arXiv:2504.07901, 2025

    2025 arXiv

  8. [16]

    Cognitive translation studies: Developments in theory and method

    Sandra L Halverson. Cognitive translation studies: Developments in theory and method. In Translation and cognition, pages 349–369. John Benjamins Publishing Company, 2010

    2010

  9. [17]

    R1-t1: Fully incentivizing translation capability in llms via reasoning learning

    Minggui He, Yilun Liu, Shimin Tao, Yuanchang Luo, Hongyong Zeng, Chang Su, Li Zhang, Hongxia Ma, Daimeng Wei, Weibin Meng, et al. R1-t1: Fully incentivizing translation capability in llms via reasoning learning. arXiv preprint arXiv:2502.19735, 2025

    2025 arXiv

  10. [18]

    Exploring human-like translation strategy with large language models

    Zhiwei He, Tian Liang, Wenxiang Jiao, Zhuosheng Zhang, Yujiu Yang, Rui Wang, Zhaopeng Tu, Shuming Shi, and Xing Wang. Exploring human-like translation strategy with large language models. Transactions of the Association for Computational Linguistics, 12:229–246, 2024. doi: 10....

    2024 doi

  11. [19]

    How good are GPT models at machine translation? A comprehensive evaluation

    Amr Hendy, Mohamed Abdelrehim, Amr Sharaf, Vikas Raunak, Mohamed Gabr, Hitokazu Matsushita, Young Jin Kim, Mohamed Afify, and Hany Hassan Awadalla. How good are GPT models at machine translation? A comprehensive evaluation. CoRR, abs/2302.09210, 2023

    2023 arXiv

  12. [20]

    Metricx-24: The google submission to the wmt 2024 metrics shared task, 2024

    Juraj Juraska, Daniel Deutsch, Mara Finkelstein, and Markus Freitag. Metricx-24: The google submission to the wmt 2024 metrics shared task, 2024. URL https://arxiv.org/abs/ 2410.03983

    2024 arXiv

  13. [21]

    Findings of the WMT24 general machine translation shared task: The LLM era is here but MT is not solved yet

    Tom Kocmi, Eleftherios Avramidis, Rachel Bawden, Ondˇrej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Markus Freitag, Thamme Gowda, Roman Grundkiewicz, Barry Haddow, Marzena Karpinska, Philipp Koehn, Benjamin Marie, Christof Monz, Kenton Murray, Masaaki Nagata, M...

    2024

  14. [22]

    Training the translator

    Paul Kussmaul. Training the translator. John Benjamins Publishing Company, 1995

    1995

  15. [23]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large lan- guage model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Syste...

    2023

  16. [24]

    Camel: Communicative agents for "mind" exploration of large language model society

    Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for "mind" exploration of large language model society. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  17. [25]

    Deepseek-v3 technical report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 11

    2024 arXiv

  18. [26]

    Llamax: Scaling linguistic horizons of LLM by enhancing translation capabilities beyond 100 languages

    Yinquan Lu, Wenhao Zhu, Lei Li, Yu Qiao, and Fei Yuan. Llamax: Scaling linguistic horizons of LLM by enhancing translation capabilities beyond 100 languages. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors, Findings of the Association for Computational Linguistic...

    2024

  19. [27]

    Introducing gpt-4.1 in the api, April 2025

    OpenAI. Introducing gpt-4.1 in the api, April 2025. URL https://openai.com/index/gpt- 4-1/. Accessed: 2025-04-23

    2025

  20. [28]

    chrf: character n-gram f-score for automatic mt evaluation

    Maja Popovi´c. chrf: character n-gram f-score for automatic mt evaluation. In Proceedings of the tenth workshop on statistical machine translation, pages 392–395, 2015

    2015

  21. [29]

    A call for clarity in reporting bleu scores

    Matt Post. A call for clarity in reporting bleu scores. arXiv preprint arXiv:1804.08771, 2018

    2018 arXiv

  22. [30]

    Scaling up cometkiwi: Unbabel-ist 2023 submission for the quality estimation shared task

    Ricardo Rei, Nuno M Guerreiro, José Pombal, Daan van Stigt, Marcos Treviso, Luisa Coheur, José GC de Souza, and André FT Martins. Scaling up cometkiwi: Unbabel-ist 2023 submission for the quality estimation shared task. arXiv preprint arXiv:2309.11925, 2023

    2023 arXiv

  23. [31]

    A cognitive scientific view on technical communication and translation: Do embodiment and situatedness really make a difference? Target

    Hanna Risku. A cognitive scientific view on technical communication and translation: Do embodiment and situatedness really make a difference? Target. International Journal of Translation Studies, 22(1):94–111, 2010

    2010

  24. [32]

    Current research in bilingualism and its implications for cognitive translation and interpreting studies

    John W Schwieter, Julia Festman, and Aline Ferreira. Current research in bilingualism and its implications for cognitive translation and interpreting studies. Linguistica Antverpiensia, New Series–Themes in Translation Studies, 19, 2020

    2020

  25. [33]

    Ilya Sutskever, Oriol Vinyals, and Quoc V . Le. Sequence to sequence learning with neural networks. In Proc. of NeurIPS, pages 3104–3112, 2014

    2014

  26. [34]

    Qwq-32b: Embracing the power of reinforcement learning, 2024

    Qwen Team. Qwq-32b: Embracing the power of reinforcement learning, 2024. URL https: //qwenlm.github.io/blog/qwq-32b/. Accessed: 2025-04-23

    2024

  27. [35]

    The mqm error typology

    The MQM Council. The mqm error typology. https://themqm.org/error-types-2/ typology/, 2025. URL https://themqm.org/error-types-2/typology/. Accessed: 2025-04-27

    2025

  28. [36]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. of NeurIPS, pages 5998–6008, 2017

    2017

  29. [37]

    Drt: Deep reasoning translation via long chain-of-thought, 2025

    Jiaan Wang, Fandong Meng, Yunlong Liang, and Jie Zhou. Drt: Deep reasoning translation via long chain-of-thought, 2025. URL https://arxiv.org/abs/2412.17498

    2025 arXiv

  30. [38]

    Deep reasoning translation via reinforcement learning

    Jiaan Wang, Fandong Meng, and Jie Zhou. Deep reasoning translation via reinforcement learning. arXiv preprint arXiv:2504.10187, 2025

    2025 arXiv

  31. [39]

    Wong, and Lidia S

    Qiang Wang, Bei Li, Tong Xiao, Jingbo Zhu, Changliang Li, Derek F. Wong, and Lidia S. Chao. Learning deep transformer models for machine translation. In Proc. of ACL, pages 1810–1822, 2019

    2019

  32. [40]

    TransAgents: Build your translation company with language agents

    Minghao Wu, Jiahao Xu, and Longyue Wang. TransAgents: Build your translation company with language agents. In Delia Irazu Hernandez Farias, Tom Hope, and Manling Li, editors, Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonst...

    2024

  33. [41]

    (perhaps) beyond human translation: Harnessing multi-agent collaboration for translating ultra-long literary texts, 2024

    Minghao Wu, Yulin Yuan, Gholamreza Haffari, and Longyue Wang. (perhaps) beyond human translation: Harnessing multi-agent collaboration for translating ultra-long literary texts, 2024. URL https://arxiv.org/abs/2405.11804

    2024 arXiv

  34. [42]

    A paradigm shift in machine translation: Boosting translation performance of large language models

    Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. A paradigm shift in machine translation: Boosting translation performance of large language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11,

    2024

  35. [43]

    A paradigm shift in machine translation: Boosting translation performance of large language models

    Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. A paradigm shift in machine translation: Boosting translation performance of large language models. In The Twelfth International Conference on Learning Representations , 2024. URL https: //openreview.net/forum?id=...

    2024

  36. [44]

    X-alma: Plug & play modules and adaptive rejection for quality translation at scale

    Haoran Xu, Kenton Murray, Philipp Koehn, Hieu Hoang, Akiko Eriguchi, and Huda Khayrallah. X-alma: Plug & play modules and adaptive rejection for quality translation at scale. arXiv preprint arXiv:2410.03115, 2024

    2024 arXiv

  37. [45]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  38. [46]

    An overview of cognitive translation studies

    Chen Yi. An overview of cognitive translation studies. Canadian Social Science, 16(5):39–43, 2020

    2020

  39. [47]

    Instruct

    Nataliia Yuhan, Mykhailo Zhylin, Solomiya Antonyuk-Kyrychenko, Nataliia Popova, and Maryna Harbar. Cognitive aspects of translation: The latest research in psycholin- guistics and cognitive science. Forum for Linguistic Studies , 6(4):316–325, Oct. 2024. doi: 10.30564/fls.v6i4...

    2024 doi

  40. [49]

    translation

    Literal Translation: Also known as direct translation or word-for-word translation, it prioritizes accurate meaning while preserving the original text’s form and content in the target language. 2.Sense-for-Sense Translation: Focuses on conveying the core meaning of the origina...

  41. [50]

    Addition: Translation includes information not present in the source

  42. [51]

    Omission: Translation is missing content from the source

  43. [52]

    Mistranslation: Translation does not accurately represent the source

  44. [53]

    Expressiveness Evaluation Criteria:

    Untranslated text: Source text has been left untranslated. Expressiveness Evaluation Criteria:

  45. [54]

    Punctuation: Incorrect punctuation (for locale or style)

  46. [55]

    Spelling: Incorrect spelling or capitalization

  47. [56]

    Grammar: Problems with grammar, other than orthography

  48. [57]

    Register: Wrong grammatical register (e.g., inappropriately informal pronouns)

  49. [58]

    Inconsistency: Internal inconsistency (not related to terminology)

  50. [59]

    Elegance Evaluation Criteria:

    Character encoding: Characters are garbled due to incorrect encoding. Elegance Evaluation Criteria:

  51. [60]

    Terminology: Terminology is either non-standard, does not fit the context, or is used inconsistently

  52. [61]

    Style: Translation has stylistic problems

  53. [62]

    Locale convention: Wrong format for addresses, currency, dates, names, telephone numbers, time expressions, or other locale-specific elements

  54. [63]

    Logical Expression: Translation lacks logical coherence or does not align with the thinking patterns and language expressions of the target language

  55. [64]

    faithfulness_score

    Other: Any other issues that might affect the elegance of the translation. ## Source Text ({source_language}): {source_text} ## Translation ({target_language}): {translation} ## Evaluation: {evaluation_result} ## Output format specification: ‘‘‘json {{ 18 "faithfulness_score":...

  56. [65]

    keyword1: keyword1 in {target_language}

  57. [66]

    translation

    phrases2: phrases2 in {target_language} Zero-shot SYSTEM_PROMPT You are a helpful assistant. USER_PROMPT Translate this from {source_language} to {target_language}: {source_language}: {source_text} {target_language}: Please generate the final translation in JSON format as foll...

  58. [2024]

    OpenReview.net, 2024. 12

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.