Pith. sign in

REVIEW 4 major objections 6 minor 54 references

Can an Old Dog Be Taught New Tricks? Taking LLMs Beyond Sentence Level Translation

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read The paper claims that large language models can be steered away from sentence-by-sentence translation toward whole-document, target-oriented reformulation by combining user specifications with retrieved examples from a comparable corpus, an

desk verdict A transparent, modest applied paper on corpus-informed whole-document LLM translation; the reformulation claim is plausible but the PAT retrieval mechanism and the evaluation base are too thin to carry the weight. read the letter →

arxiv 2607.14040 v1 pith:R35RYRU6 submitted 2026-07-15 cs.CL

Alaina Brandt This is my paper
classification cs.CL
keywords whole-documenttranslationcorpus-informedpragmaticequivalencelargelanguagemodelsretrieval-augmentedgenerationreformulationqualityevaluationEnglish-Spanish
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

This paper asks whether large language models can be moved beyond the sentence-by-sentence paradigm that dominates automatic translation, toward whole-document reformulation that fits the target language's discourse and stylistic norms. It presents PAT, a retrieval-augmented system that pairs user-configured specifications with examples retrieved from a comparable corpus of authentic long-form English and Spanish writing, then generates the whole document in one pass. Evaluations of six English-to-Spanish translations show that a bare "translate this" prompt yields close, phrase-following output, while specifications- and corpus-informed prompts produce observable reformulation: paragraphs split or merged, ideas reordered, source repetition dropped, and idiomatic Spanish phrasing introduced. The paper's central claim is that LLMs can be moved toward reformulation, though the reformulations are not yet consistently effective, and the open work is making that reformulation serve quality.

What carries the argument

The carrying mechanism is PAT's system prompt, which combines user-configured specifications (subject field, language variants, text type, purpose, audience, perspective, distribution scope) with retrieved corpus context, and instructs the model to read and comprehend the full source text before generating. Retrieval operates on a comparable corpus of authentic long-form texts indexed at paragraph, section, and document levels; a document-level embedding of the source is compared against corpus embeddings, and the user selects which retrieved passages enter the prompt. The design assumption doing heavy lifting is that document-level embeddings encode stylistic and discourse similarity alongs

What would settle it

Run PAT twice on the same source with identical specifications, once with retrieved passages that are topically matched but stylistically opposite (for example, terse news prose retrieved for an ornate essay source) and once with stylistically matched passages. If the two outputs are indistinguishable in reformulation style, then retrieval is not carrying stylistic information and the embedding-stylistic assumption fails. A simpler version: replace retrieved passages with randomly drawn corpus passages; if output quality and reformulation do not change, the corpus-context mechanism is inert.

Watch

Extended reading notes

Core claim

The central claim is that LLMs can be moved toward reformulation and away from the sentence-by-sentence paradigm, but that the reformulations are not yet consistently effective. Using PAT, the authors demonstrate that when an LLM is given user-configured specifications and retrieved paragraph-, section-, and document-level examples from a comparable corpus of authentic long-form English and Spanish writing, its output shows observable whole-document reformulation: paragraphs split and consolidated, ideas reordered, source repetition omitted, and target-idiomatic phrasing introduced. A limited prompt produced no meaningful reformulation across all tested texts. The authors are careful to call

Load-bearing premise

The load-bearing premise, which the paper itself flags as a design assumption in Section 4.2, is that document-level embeddings capture stylistic and discourse similarity alongside topical similarity when retrieving corpus examples; if embeddings only match topic, the retrieved passages would be topically relevant but stylistically useless or misleading.

Editorial extensions

If this is right

  • A bare 'translate this text' prompt is not a neutral baseline; it actively selects sentence-level behavior, so prompt wording becomes a controllable variable in translation-system design.
  • Specifications alone can induce meaningful reformulation, meaning that even without a corpus, setting audience, purpose, and text type in the prompt moves output away from close transfer.
  • Retrieved corpus passages can transfer target-language idiom and discourse organization into the output, so corpus construction (authentic, comparable, well-formatted examples) becomes a direct quality lever.
  • Excess or poorly delimited context can dilute reformulation, so retrieval quantity and formatting are parameters to tune, not just content.
  • Since PAT outputs are unverified drafts, the practical deployment path is a human-in-the-loop workflow where a professional translator verifies and adjusts the reformulated draft.

Reading between the lines

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

  • Editorial inference: if the embedding-stylistic assumption is validated, retrieval could be directed by style similarity alone, turning the system into an explicit style-transfer device: 'translate this into the register of these example texts.'
  • Editorial inference: the observed evaluator preference for less reformulated output, if it reflects industry training rather than reader experience, implies that current quality-evaluation conventions may systematically undervalue pragmatic reformulation; a comparative study of professional-translator versus lay-reader judgments would test this.
  • Editorial inference: the finding that the word 'translate' itself biases output toward sentence-level processing suggests a broader principle: LLM behavior on generation tasks is steered by task vocabulary, and rewording the task label may be as powerful as changing the model.
  • Editorial inference: the corpus-informed condition's style transfer worked when context was limited and well delimited; a testable extension is to vary context length and formatting systematically to find the point at which additional examples stop helping and start acting as noise.
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

4 major / 6 minor

Summary. The paper argues that LLM translation need not remain sentence-by-sentence. It first reports (Section 3) that a bare 'translate this text' prompt yields close sentence-level output across four models, while a reformulation prompt yields freer output in 13/20 cases. It then presents PAT, a RAG pipeline that combines user-configured specifications with retrieved paragraph/section/document passages from a comparable English/Spanish corpus, and uses Gemini 2.5 Pro for whole-document generation. Six production translations across three projects were evaluated by two MQM-trained annotators with a customized typology. The paper reports that limited prompts produce no meaningful reformulation; specifications alone produce minor-to-substantial reformulation; and specifications-plus-corpus produce moderate-to-substantial reformulation, though quality evaluations did not consistently favor the more reformulated texts. The central claim is that LLMs can be moved toward reformulation and away from sentence-level translation, with effectiveness still requiring future work.

Significance. If the result holds, the paper is a useful existence proof for a human-in-the-loop, corpus-informed translation workflow, and it provides a concrete architecture (PAT) plus an evaluation framework (customized MQM) that others can adapt. The manuscript is commendably candid: it labels its embedding style-sensitivity a design assumption (footnote 2), reports low interannotator agreement (Table 4), and makes its system prompt, corpora, and TQE datasets available (Annex 1). These strengthen reproducibility. The main limitations—untested retrieval assumption, informal reformulation classification, and reliability of quality judgments—are all addressable and do not undermine the qualitative existence proof, but they do prevent the corpus mechanism and effectiveness claims from being fully supported.

major comments (4)
  1. [§4.2, footnote 2] Footnote 2 states that document-level embeddings encode stylistic alongside propositional similarity is a design assumption, but no retrieval evaluation is reported. Section 6.2's evidence consists of the author's selected examples ('illustrated relatively shorter, more direct sentence structures'), not a systematic comparison. Because the Context condition also contains the full PAT system prompt, the extra reformulation seen in TQE-6 Text 1 (Table 2) could be due to the prompt or to any authentic target-language passage, not necessarily to style-sensitive retrieval. Add a control (e.g., random or topic-only retrieved passages) or a blind style-relevance judgment of retrieved versus random passages, or explicitly downgrade the corpus-retrieval claim.
  2. [§3, Table 1; §4.4] The 'all twenty' close-translation result and the 13/20 freer count in Table 1 are used to motivate the paper and to choose Gemini 2.5 Pro as PAT's engine (§4.4). No example outputs, coding rubric, or second annotator are given for this classification, and the forty translations are not listed in the Annex 1 repository. If the classification is unreliable or the model is selected on the outcome, the production-phase reformulation may be inflated. Provide the outputs/rubric with agreement, or recast Section 3 as pilot evidence and do not make model selection depend on it.
  3. [§6.1, Table 2] The degree-of-reformulation scale (None/Minor/Moderate/Substantial) is the paper's primary outcome for the central claim, yet it has no operational definitions (e.g., number of sentence-boundary changes, paragraph reorganization, additions/omissions), no statement of who assigned the ratings, and no reliability check. Table 3 illustrates only a few sentences in one project. Add a rubric and at least two independent raters for the Table 2 classifications, or present them clearly as the author's qualitative impressions rather than as measurements.
  4. [§6.3–6.4, Tables 4–5] Interannotator span agreement is 49.5–61.4% and category agreement falls below the paper's own 60% target; only two evaluators contribute to the production-phase scores. The 'not always to effect' component of the claim rests on these quality comparisons (e.g., TQE-6 Context: 51 spans, mean correspondence 2.5 vs. Prompt: 31 spans, 3.0). With agreement near chance on error categories, these differences should be labeled exploratory, or a third evaluator or holistic protocol with demonstrated reliability should be added before drawing effectiveness conclusions.
minor comments (6)
  1. [§3, Table 1] The note that only ten of forty translations reflected expected English-to-Spanish expansion is unexplained; state how expansion was measured and why it matters for reformulation.
  2. [§4.4] The causal explanation contrasting DeepSeek's and Gemini's training objectives is speculative; either support with model documentation or mark it explicitly as a hypothesis.
  3. [§6.3, Fig. 1] Figure 1 is referenced but does not appear in the manuscript; ensure the figure is included and legible in the final version.
  4. [Table 7] 'torriente de conciencia' appears to be a typo for 'torrente/corriente de conciencia'.
  5. [References] Reference [17] lists 'Salama-C.' in incomplete form; should be 'Salama-Carr, M.'
  6. [§6.2] The statement that the English section in Project 7 was chosen because the source genre has no Spanish equivalent is asserted rather than demonstrated; a brief justification would help.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the reformulation findings rest on disclosed empirical comparisons, with one non-load-bearing self-citation.

full rationale

This is an empirical systems/quality-evaluation paper; there is no equation-level derivation whose outputs could reduce to inputs by construction. The central claim—that prompts, specifications, and corpus context can move LLM output toward reformulation—rests on controlled comparisons: Section 3's limited-prompt versus reformulation-prompt translations across four models, and Section 6's prompt/context comparisons. The one place where a prior choice could inflate the result is Section 4.4: 'Gemini 2.5 Pro was selected as PAT’s translation engine on the basis of its receptiveness to whole-document reformulation (cf. Table 1).' This is a disclosed selection on an outcome measured in the same paper; it limits the strength of the Section 6 observations as independent evidence, but it is not a fitted parameter renamed as a prediction, and Section 3 already establishes the basic capability independently. Footnote 2 explicitly states that the stylistic-embedding assumption is a design assumption 'positioned to test this assumption directly,' so no result is presented as derived from it. The only self-citation, reference [6], supports a background definition of specifications and is not load-bearing. Therefore no circular step is identifiable; the score reflects one minor non-load-bearing self-citation rather than any reduction of the central claim to its inputs.

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

The claims rest less on mathematical assumptions than on domain and methodological premises: that comparable-corpus examples retrieved by embeddings can shape style, that reformulation is desirable for pragmatic equivalence, and that two trained MQM evaluators provide a valid quality signal. No numeric free parameters are fitted; the system parameters (retrieved-passage counts, Gemini defaults) are design choices rather than fitted values.

assumptions (5)
  • domain assumption Document-level embeddings encode stylistic along with propositional similarity.
    Footnote 2 in §4.2 states this is a design assumption; retrieval relevance depends on it, and the paper says it is positioned to test it directly but does not test it here.
  • domain assumption English and Spanish differ systematically in discourse organization and rhetorical conventions such that reformulation is necessary for good translation.
    §2 draws on contrastive rhetoric sources [24,25]; the entire motivation for PAT rests on this.
  • domain assumption A professional translator's workflow - specifications plus research - can be approximated by prompts plus retrieved corpus context.
    §1 and §4.3; the paper's system design assumes this equivalence.
  • domain assumption LLMs default to sentence-by-sentence translation without explicit guidance.
    §3 asserts this based on an informal demonstration; it is the premise for why PAT needs its prompt.
  • domain assumption MQM-based annotations by two trained evaluators are a valid measure of translation quality.
    §5; validity is assumed despite agreement below the paper's own 60% target for category-level errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can an Old Dog Be Taught New Tricks? Taking LLMs Beyond Sentence Level Translation." pith.science (2026). https://pith.science/paper/R35RYRU6

@misc{pith2026260714040,
  author       = {Pith},
  title        = {Pith review of: Can an Old Dog Be Taught New Tricks? Taking LLMs Beyond Sentence Level Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R35RYRU6}},
  note         = {Machine review of arXiv:2607.14040}
}
read the original abstract

Automatic translation systems, from CAT tools to MT, overwhelmingly treat translation as a sentence-by-sentence act. This paper asks whether LLMs can be moved beyond that paradigm through whole-document, corpus-informed translation. We present PAT (Pragmatic Auto-Translator), a RAG-based system that pairs user-configured specifications with context from a comparable corpus of authentic longform texts in U.S. English and Latin American Spanish, passing retrieved paragraph-, section-, and document-level examples to an LLM for whole-document generation. The goal is draft translation for professional verification: target texts reformulated to fit their Spanish-language context, where discourse organization, rhetorical style, and pragmatic norms differ meaningfully from English. We evaluated six automatic translations of essays on generative AI across three projects using a customized MQM typology, assessed by two trained evaluators working from U.S. English into LATAM and Mexican Spanish. Results show that a limited prompt produced no meaningful reformulation, and specifications and corpus-informed translations at times showed substantial reformulation, though not always to effect. We find that LLMs can be moved toward reformulation and away from the sentence-by-sentence paradigm, though more work is needed to improve the effectiveness of those reformulations. In this paper, we discuss considerations related to automatic translation system design, corpus construction, and translation quality evaluation methodology and results.

Figures

Figures reproduced from arXiv: 2607.14040 by the authors.

Figure 1
Figure 1. Partial span agreement (F1, character-level) and category agreement as a proportion of partial span matches, measured between two evaluators across seven TQE projects in chronological order. Training-phase projects (TQE-1 through TQE-4) include into-English and into-Spanish directions; production-phase projects (TQE-5 through TQE-7) are exclusively into￾Spanish. The upward trend in partial span agreement across the … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 4 linked inside Pith

  1. [1]

    West Conshohocken, PA, USA (2025)

    ASTM International: ASTM F2575-25 Standard Practice for Language Translation. West Conshohocken, PA, USA (2025)

    2025

  2. [2]

    Routledge, London (2014)

    Reiß, K., Vermeer, H.J.: Towards a general theory of translational action: Skopos theory explained. Routledge, London (2014)

    2014

  3. [3]

    arXiv preprint arXiv:2304.12959 (2024)

    Post, M., Junczys-Dowmunt, M.: Escaping the sentence-level paradigm in machine translation. arXiv preprint arXiv:2304.12959 (2024)

    arXiv 2024

  4. [4]

    MQM (Multidimensional Quality Metrics), https://themqm.org, last accessed 2026/06/24

    2026

  5. [5]

    J.: The industrialization of translation: Causes, consequences and challenges

    Dunne, K. J.: The industrialization of translation: Causes, consequences and challenges. In: Translation Spaces, 1(1), 143-168 (2012)

    2012

  6. [6]

    A.: Translation and Localization Project and Process Managers

    Brandt. A.: Translation and Localization Project and Process Managers. In: Massey G., Ehrensberger-Dow, M. and Angelone, E (eds). Handbook of the Language Industry. Walter de Gruyter GmbH, Berlin/Boston (2024)

    2024

  7. [7]

    In: Perspectives, 25(3), 464–477 (2017)

    Moorkens, J.: Under pressure: translation in times of austerity. In: Perspectives, 25(3), 464–477 (2017)

    2017

  8. [8]

    Translation Spaces, 9(1) (2020)

    do Carmo, F.: ‘Time is Money’ and the value of translation. Translation Spaces, 9(1) (2020)

    2020

Show all 54 references
  1. [9]

    Results of an ethnographic study in three translation services and agencies

    LeBlanc, M.: Translators on translation memory (TM). Results of an ethnographic study in three translation services and agencies. In: The International Journal for Translation & Interpreting Research, 5(2) (2023)

    2023

  2. [10]

    Geneva, Switzerland (2019)

    International Organization for Standardization: ISO 30042:2019 Management of terminology resources — TermBase eXchange (TBX). Geneva, Switzerland (2019)

    2019

  3. [11]

    Geneva, Switzerland (2024)

    International Organization for Standardization: ISO 21720:2024 XLIFF (XML Localization Interchange File Format). Geneva, Switzerland (2024)

    2024

  4. [12]

    TMX, https://okapiframework.org/wiki/index.php/TMX, last accessed 2026/06/11 15

    2026

  5. [13]

    and Artzi, Y.: BERTScore: Evaluating Text Generation with BERT

    Zhang, T., Kishore, V., Wu, F., Weinberger, K. and Artzi, Y.: BERTScore: Evaluating Text Generation with BERT. In: International Conference on Learning Representations (2020)

    2020

  6. [14]

    In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, 2685–2702 (2020)

    Rei, R., Stewart, C., Farinha, A., Lavie, A.: COMET: A Neural Framework for MT Evaluation. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, 2685–2702 (2020)

    2020

  7. [15]

    Shared Task: Automated Translation Quality Evaluation Systems, https://www2.statmt.org/wmt26/mteval-task.html, last accessed 2026/06/11

    2026

  8. [16]

    Unpublished lecture notes, UW-Milwaukee (2012)

    Bolduc, M.: Translation Theory in History (I). Unpublished lecture notes, UW-Milwaukee (2012)

    2012

  9. [17]

    In: Baker, M

    Salama-C.: French tradition. In: Baker, M. and Saldanha, G. (eds). Routledge Encyclopedia of Translation Studies. 2nd edn. Routledge, London and New York, pp. 404-410 (2011)

    2011

  10. [18]

    T.: The Best Kind of Orator

    Cicero, M. T.: The Best Kind of Orator. (46 B.C.E.) Translator: Hubbell, H.M. In: Robinson, D. Western Translation Theory from Herodotus to Nietzsche. St. Jerome Publishing, Manchester (1997)

    1997

  11. [19]

    In: Venuti, E

    Nida, E.: Principles of correspondence. In: Venuti, E. The Translation Studies Reader. 5th edn. Routledge, London (2026)

    2026

  12. [20]

    Routledge, London (1995)

    Venuti, L.: The Translator’s Invisibility: A History of Translation. Routledge, London (1995)

    1995

  13. [21]

    Bloomsbury Academic, London & New York (2023)

    Morini, M: The Pragmatic Translator. Bloomsbury Academic, London & New York (2023)

    2023

  14. [22]

    In: In Other Words: A coursebook on translation

    Baker, M.: Pragmatic Equivalence. In: In Other Words: A coursebook on translation. 2nd edn. Routledge, London and New York, pp. 230–273 (2011)

    2011

  15. [23]

    In: Che, W., Nabende, J., Shutova, E., Pilehvar, M.T

    Guo, Y., Conia, S., Zhou, Z., Li, M., Potdar, S., Xiao, H.: Do Large Language Models have an English Accent? Evaluating and Improving the Naturalness of Multilingual LLMs. In: Che, W., Nabende, J., Shutova, E., Pilehvar, M.T. (eds.) Proceedings of the 63rd Annual Meeting of th...

    2025

  16. [24]

    R.: Discourse Features of Written Mexican Spanish: Current Research in Contrastive Rhetoric and Its Implications

    Montaño-Harmon, M. R.: Discourse Features of Written Mexican Spanish: Current Research in Contrastive Rhetoric and Its Implications. Hispania, 74(2), 417–425 (1991)

    1991

  17. [25]

    In: Santiago Guervós, F.J

    Sánchez Jiménez, D.: Estudios transculturales del texto escrito en español: investigaciones y resultados. In: Santiago Guervós, F.J. de, Bongaerts, H., Sánchez Iglesias, J.J., Seseña Gómez, M. (eds.) Del texto a la lengua: la aplicación de los textos a la enseñanza-aprendizaje...

    2011

  18. [26]

    In: Connor, U., Nagelhout, E., Rozycki, W

    Pak, C., Acevedo, R.: Spanish language newspaper editorials from Mexico, Spain, and the U.S.. In: Connor, U., Nagelhout, E., Rozycki, W. (eds.), Contrastive Rhetoric: Reaching to intercultural rhetoric, pp. 123-145. John Benjamins Publishing Company, (2008)

    2008

  19. [27]

    In: Bouamor, H., Pino, J., Bali, K

    Wang, L., Lyu, C., Ji, T., Zhang, Z., Yu, D., Shi, S., Tu, Z.: Document-Level Machine Translation with Large Language Models. In: Bouamor, H., Pino, J., Bali, K. (eds.) Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 16646–16661. Ass...

    2023

  20. [28]

    In: Ebrahimi, A., Haider, S., Liu, E., Haider, S., Pacheco, M.L., Wein, S

    Sun, Y., Zhu, D., Chen, Y., Xiao, E., Chen, X., Shen, X.: Fine-Grained and Multi-Dimensional Metrics for Document- Level Machine Translation. In: Ebrahimi, A., Haider, S., Liu, E., Haider, S., Pacheco, M.L., Wein, S. (eds.) Proceedings of the 2025 Conference of the Nations of ...

    2025

  21. [29]

    In: Somers, H

    Knowles, F.: Lexical cartography in LSP texts. In: Somers, H. (ed.) Terminology, LSP and Translation: Studies in language engineering in honour of Juan C. Sager, pp. 125–140. John Benjamins, Amsterdam (1996)

    1996

  22. [30]

    In: Massey, G., Ehrensberger-Dow, M., Angelone, E

    Warburton, K.: Terminology Managers. In: Massey, G., Ehrensberger-Dow, M., Angelone, E. (eds.) Handbook of the Language Industry: Contexts, Resources and Profiles, pp. 179–201. De Gruyter Mouton, Berlin (2024)

    2024

  23. [31]

    In: Somers, H

    Baker, M.: Corpus-based translation studies: The challenges that lie ahead. In: Somers, H. (ed) Terminology, LSP and Translation: Studies in language engineering in honour of Juan C. Sager, pp. 175–186. John Benjamins, Amsterdam (1996)

    1996

  24. [32]

    Bender, E.M., Gebru, T., McMillian-Major, A., Shmitchell, S.: On the Dangers of Stochastic Parrots: Can Language Models be Too Big? In: Conference on Fairness, Accountability, and Transparency (FAccT ’21), ACM, New York, NY, USA (2021)

    2021

  25. [33]

    Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J.W., Wallach, H., Daumé III, H., Crawford, K.: Datasheets for datasets. Commun. ACM 64(12), 86–92 (2021)

    2021

  26. [34]

    In: The World Wide Web Conference, WWW ’19, pp

    Jiang, J.Y., Zhang, M., Li, C., Bendersky, M., Golbandi, N., Najork, M.: Semantic text matching for long-form documents. In: The World Wide Web Conference, WWW ’19, pp. 795–806. Association for Computing Machinery, New York, NY, USA (2019)

    2019

  27. [35]

    Jina Embeddings v3: A Frontier Multilingual Embedding Model, https://jina.ai/news/jina-embeddings-v3-a-frontier- multilingual-embedding-model/, last accessed 2026/06/26 16

    2026

  28. [36]

    jina-embeddings-v3, https://jina.ai/models/jina-embeddings-v3, last accessed 2026/06/25

    2026

  29. [37]

    FacebookAI/xlm-roberta-large, https://huggingface.co/FacebookAI/xlm-roberta-large, last accessed 2026/06/25

    2026

  30. [38]

    MTEB Multilingual v2, https://huggingface.co/spaces/mteb/leaderboard, last accessed 2026/06/25

    2026

  31. [39]

    jina-embeddings-v5-text-small, https://jina.ai/models/jina-embeddings-v5-text-small, last accessed 2026/06/25

    2026

  32. [40]

    arXiv:2505.09388 (2025)

    Yang, A., et al.: Qwen3 Technical Report. arXiv:2505.09388 (2025)

    2025 arXiv

  33. [41]

    In: Proceedings of the Tenth Conference on Machine Translation, pp

    Kayano, Y., Sugawara, S.: Specification-aware machine translation and evaluation for purpose alignment. In: Proceedings of the Tenth Conference on Machine Translation, pp. 113–141. Association for Computational Linguistics (2025)

    2025

  34. [42]

    To whom it may concern

    Merx, R., Vylomova, E., Cohn, T.: Beyond "To whom it may concern": Tailoring Machine Translation to Audience and Intent. arXiv:2606.03259 (2026)

    2026 arXiv

  35. [43]

    Standards-based structured translation specifications, https://www.tranquality.info/specs/, last accessed 2026/06/25

    2026

  36. [44]

    Geneva, Switzerland (2024)

    International Organization for Standardization: ISO 11669:2024 Translation projects — General guidance. Geneva, Switzerland (2024)

    2024

  37. [45]

    arXiv:2606.19348 (2026)

    DeepSeek-AI: DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence. arXiv:2606.19348 (2026)

    2026

  38. [46]

    arXiv:2507.06261 (2025)

    Gemini Team, Google: Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities. arXiv:2507.06261 (2025)

    2025 arXiv

  39. [47]

    Gemini 2.5 Pro, https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/2-5-pro, last accessed 2026/06/25

    2026

  40. [48]

    Label Studio, https://labelstud.io, last accessed 2026/06/25

    2026

  41. [49]

    The MQM CORE Typology, https://themqm.org/the-mqm-typology/, last accessed 2026/06/26

    2026

  42. [50]

    ASTM WK54884 New Practice for Holistic Human and AI-based Quality Evaluation System for Translation, https://www.astm.org/membership-participation/technical-committees/workitems/workitem-wk54884, last accessed 2026/06/26

    2026

  43. [51]

    Decision Tree, https://themqm.org/error-types-2/decisiontree/, last accessed 2026/06/26

    2026

  44. [52]

    Journal of the American Medical Informatics Association 12(3), 296-298 (2005)

    Hripcsak G., Rothschild A.S.: Agreement, the f-measure, and reliability in information retrieval. Journal of the American Medical Informatics Association 12(3), 296-298 (2005)

    2005

  45. [53]

    Los Angeles Review of Books, https://lareviewofbooks.org/blog/provocations/artificial- intelligence/, last accessed 2026/06/25 (2020)

    Klonik, K.: What AI Is Not. Los Angeles Review of Books, https://lareviewofbooks.org/blog/provocations/artificial- intelligence/, last accessed 2026/06/25 (2020)

    2026

  46. [54]

    Hall, L.: I Stumbled Across My Boyfriend’s ChatGPT and It Ended Our Relationship, https://lindseyhallwrites.substack.com/p/i-read-my-boyfriends-chatgpt-and, last accessed 2026/07/09 Annexes Annex 1. This paper makes reference to various external resources related to this resea...

    2026

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.