Pith. sign in

REVIEW 3 major objections 4 minor 67 references

This paper claims that a 2.6B latent-reasoning looped language model, adapted with per-pair LoRA adapters, matches or exceeds translation systems three to five times larger across 32 directions.

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

2026-08-01 14:47 UTC pith:JPNHJ4UP

load-bearing objection Believable saturation story, but the low-resource SOTA claim rests on a single seedless u=4 run—needs error bars and a depth sweep before that headline stands. the 3 major comments →

arxiv 2607.18618 v1 pith:JPNHJ4UP submitted 2026-07-21 cs.CL cs.AIcs.LG

LatentMT: Machine Translation with Latent Reasoning

Wei-Rui Chen , Samar M. Magdy , Chiyu Zhang , Wenhui Zhu , Zhipeng Wang , Muhammad Abdul-Mageed This is my paper
classification cs.CL cs.AIcs.LG
keywords latent reasoninglooped language modelsmachine translationLoRA adaptationrecurrent depthparameter-efficient fine-tuninglow-resource translationefficiency-quality tradeoff
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.

LatentMT tests whether a machine translation system can trade parameter count for repeated computation inside a small shared backbone. The paper adapts a frozen 2.6-billion-parameter latent-reasoning looped language model with lightweight per-language-pair LoRA adapters, and reports that this combination is competitive in high-resource translation, state-of-the-art on mid- and low-resource benchmarks, and comparable overall to models three to five times larger. A second claim is that recurrent depth is a quality dial: translation improves sharply from one to roughly four recurrent steps and then saturates. The why-care is practical: if a small model can match much larger systems by unrolling the same weights a few extra times, strong translation becomes feasible on local or edge hardware at lower training and inference cost.

Core claim

On its own terms, the paper's discovery is that latent-reasoning looped language models are a viable backbone for machine translation. The authors freeze a 2.6B-parameter looped model, attach separate LoRA adapters for each language pair, and train only those adapters. Across 32 directions spanning high-, mid-, and low-resource languages, the resulting LatentMT system reaches quality comparable to systems three to five times larger, and it beats both an 8B same-protocol baseline and a state-of-the-art 8B multilingual model on average for low-resource languages. The paper also claims that the relationship between recurrent depth and quality is an early-gain-then-saturation curve, supported by

What carries the argument

The load-bearing mechanism is the looped language model (LoopLM): a shared Transformer stack applied u times to the hidden state before each next-token prediction, so extra computation costs no extra parameters. LatentMT couples this with per-language-pair LoRA adapters on a frozen backbone, and treats u as a fixed experimental variable. The saturation analysis uses two observables — cosine distance between last-hidden-layer states at adjacent recurrent steps and the rate at which the top-1 predicted token flips between steps — to show that later recurrent transitions change less.

Load-bearing premise

The low-resource 'state-of-the-art' claim rests on a randomly sampled 30-language subset and a fixed recurrent depth of u=4 for which no seed or sweep is reported; if that subset or depth was selected after inspecting test-set scores, the result could be a post-selection artifact rather than a real advantage.

What would settle it

Re-run the 30-language low-resource evaluation with a pre-registered seed, a held-out language list, and recurrent depth chosen on development data rather than test data; if the 2.6B model no longer beats the 8B baselines on average BLEU or chrF++, the central quality claim for low-resource translation is refuted.

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

If this is right

  • A 2.6B latent-reasoning model with lightweight adapters can reach translation quality comparable to systems three to five times larger, shifting the scaling axis from parameters to per-token recurrent compute.
  • Recurrent depth functions as a controllable cost-quality dial: going from u=1 to u=4 yields the main gains, and u=2 retains roughly 98% of peak quality while cutting training compute and inference latency relative to a larger baseline.
  • Because the backbone stays frozen and adapters are swapped per language pair, adding or changing translation directions does not require retraining the full model.
  • Translation quality saturates quickly because the hidden-state trajectory stabilizes, so the same early-gain pattern should be expected when applying latent reasoning to other generation tasks.
  • The reported memory and latency reductions make local or edge deployment of competitive translation more plausible for sensitive or low-connectivity settings.

Where Pith is reading between the lines

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

  • Going beyond the paper: the knee at u=2 suggests an adaptive depth scheduler — using one or two steps for easy text and four for hard text — could preserve most quality at a fraction of the cost; this is not tested in the paper.
  • Going beyond the paper: the frozen-backbone-plus-adapters design implies the system can be extended to new language pairs without full retraining, but also raises an open question of whether adapter interference appears beyond 32 directions; the paper does not address it.
  • Going beyond the paper: because the mechanistic saturation is measured on translation only, one testable extension is to compare LoRA-adapted looped models on summarization or code generation to see whether the early-gain curve is task-dependent.
  • Going beyond the paper: the efficiency comparison uses a reasoning baseline that emits explicit chain-of-thought tokens; a cleaner control would be a non-looped 2.6B model with the same LoRA adaptation, which would isolate whether the gains come from latent recurrence or from the adapter training itself.

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

3 major / 4 minor

Summary. The paper introduces LatentMT, a machine-translation approach that combines a frozen 2.6B-parameter looped language model (Ouro-2.6B-Thinking) with per-language-pair LoRA adapters. It systematically varies the fixed recurrent depth u from 1 to 6 and evaluates on three benchmark families: DRT (English–Chinese, high-resource), ArzEn-MultiGenre (English–Egyptian Arabic, mid-resource), and a 30-language OLDI/Flores subset (low-resource). The central empirical findings are that translation quality improves sharply from u=1 to roughly u=4 and then saturates; that this pattern is mirrored by human error annotation and by mechanistic measures of hidden-state change (cosine distance and top-1 flip rate); and that LatentMT is more parameter- and compute-efficient than a comparable Qwen3-8B SFT baseline. The paper claims competitiveness with models three to five times larger and state-of-the-art results on mid- and low-resource languages.

Significance. If the results hold, this is a useful and timely contribution: it provides the first systematic study of latent-reasoning looped LMs for MT, treats recurrent depth as a controlled experimental variable, and combines automatic metrics with human error analysis and mechanistic diagnostics. The controlled u-sweep on DRT and ArzEn is a real strength, as is the same-protocol comparison with Qwen3-8B SFT. The efficiency analysis is also valuable. However, the headline 'state-of-the-art on low-resource languages' claim is currently supported by a single unreported configuration on a seedless 30-language subset with no error bars, and the comparison set is narrow. The paper is worth publishing after these load-bearing issues are addressed.

major comments (3)
  1. [§4.3, Table 4] The low-resource SOTA claim is carried by a single configuration, LatentMT-2.6B4, evaluated on a 30-language OLDI/Flores subset described only as 'randomly sampled' (§3.2) with no seed. No recurrent-depth sweep, no repeated subsamples, and no variance or confidence intervals are reported for this benchmark. The average margins over EMMA-500-8B are 0.20 BLEU and 0.81 chrF++, while EMMA-500 wins on roughly half of the 30 rows; the positive average is driven by a few favorable languages. Without a seed, a sweep, and error bars, the possibility that the subset or the u=4 choice was selected after inspecting test-set scores cannot be ruled out. Please report the subset seed, a u sweep (at least u=2, 4, 6), bootstrap CIs or paired significance tests, and ideally repeated random subsets.
  2. [Abstract; §4.3] The 'state-of-the-art on low-resource languages' claim is benchmarked against only two systems: Qwen3-8B SFT and EMMA-500-8B. Strong multilingual MT baselines such as NLLB-200, M2M-100, or MADLAD-400 are absent, so the SOTA claim is not established. The DRT and ArzEn comparisons are also over a small set of reference systems. Either broaden the comparison set or temper the claim to 'competitive with the compared systems.'
  3. [§3.2, Table 3, footnote 11] All experiments are single-run, and some evaluation splits are randomly generated rather than inherited: the ArzEn train/validation/test partition is described in footnote 11 as randomly sampled to match reported sizes, but no seed is given, and the same applies to the OLDI subset. Consequently, all numeric comparisons (e.g., LatentMT4 vs. LatentMT5 in Tables 2 and 3, and the u=4 peak) lack variance estimates, and the results cannot be independently reproduced. Please run multiple seeds or provide at least one alternative split, and report standard deviations or confidence intervals for the headline numbers.
minor comments (4)
  1. [Global / reproducibility] No code, model checkpoints, or training/evaluation seeds are provided. Add a reproducibility statement with seeds, scripts, and hyperparameters for the Qwen3-8B SFT baseline.
  2. [§3.2] Typo: 'exeamplares' should be 'examples' in the Evaluation paragraph. Also, the Qwen3-8B SFT protocol is not fully specified (LoRA vs. full fine-tuning, epochs, learning rate), despite being the same-protocol control.
  3. [§5.2 / Figure 4] The efficiency analysis reports only relative reductions. Include absolute training time, GPU-hours, memory, and inference latency numbers so readers can verify the 43–77% compute savings. Figure 4 lacks error bars and axis units.
  4. [Appendix A] The human error analysis uses 30/60 segments and a single annotator with no inter-annotator agreement measure. Report the annotation guidelines and, ideally, a second annotator or an IAA score for at least a subset.

Circularity Check

0 steps flagged

No significant circularity: all headline results are empirical measurements on external benchmarks; one non-load-bearing self-citation (LQM) and subset-selection transparency issues do not make the derivation circular.

full rationale

LatentMT's claims are empirical rather than derived-from-inputs. Translation quality (Tables 2–4) is measured by evaluating trained checkpoints on held-out DRT, ArzEn-MultiGenre, and OLDI/Flores test splits with external metrics (BLEU, chrF++, COMET, COMETKiwi); no equation in the paper predicts these scores from the training objective or from a fitted parameter renamed as a prediction. The recurrent-depth saturation result is supported by automatic metrics plus independent mechanistic measurements (Eqs. 7–9), and the human error analysis in Appendix A is auxiliary; it uses the LQM taxonomy, whose authors overlap with this paper (Magdy et al., 2026), but that self-citation is evaluation tooling, not the load-bearing evidence for the saturation claim. The OLDI/Flores low-resource result rests on a single u=4 run over a 'randomly sampled' 30-language subset with no seed or error bars, and footnote 11 discloses that ArzEn splits were randomly sampled because the original release omits partitions; these are transparency/statistical-selection limitations, not definitional circularity. No step in the derivation chain reduces to its own input by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

Central claim rests on the pretrained LoopLM backbone, the validity of automatic/linguistic metrics, and design choices (LoRA ranks, fixed u). The mechanistic saturation analysis adds a new interpretation of hidden-state convergence, but it is not an independent benchmark.

free parameters (2)
  • fixed recurrent depth u = u=4 for OLDI/Flores; u=5 quality-optimal and u=2 knee for DRT
    u is swept in high/mid-resource tables but only u=4 is reported for low-resource; if chosen by test performance it is a fitted parameter affecting SOTA claim.
  • LoRA ranks/alphas per dataset = rank 32/alpha 64 for DRT and ArzEn; rank 16/alpha 32 for OLDI/Flores
    Manual hyperparameters; not swept, likely not critical, but listed for completeness.
axioms (5)
  • domain assumption Ouro-2.6B-Thinking is a valid latent-reasoning LoopLM, and its frozen weights implement recurring computation as described.
    The paper relies on the pretrained backbone's behavior; no independent verification is provided beyond the model card.
  • domain assumption COMET, COMETKiwi, BLEU, chrF++ reflect translation quality; differences of ~1 BLEU are meaningful.
    Metrics are used to compare systems without confidence intervals or significance tests.
  • domain assumption Fixed-u training generalizes to fixed-u inference for the same u.
    The training objective (Eq 6) uses the same P^u as inference; no evidence that training at one u transfers to another.
  • ad hoc to paper Cosine distance and top-1 flip rate across recurrent steps capture when additional computation stops adding information.
    These measures are introduced by the paper as mechanistic evidence; they are descriptive and not established predictors of translation quality.
  • domain assumption The OLDI/Flores 30 languages are a random sample, and training on OLDI-seed does not overlap with Flores-200 test.
    No seed or leakage analysis is provided; a train/test overlap or non-random subset would inflate low-resource scores.

pith-pipeline@v1.3.0-alltime-deepseek · 14568 in / 12834 out tokens · 106117 ms · 2026-08-01T14:47:52.294591+00:00 · methodology

0 comments
read the original abstract

Latent-reasoning looped language models (LoopLMs) offer a different scaling path for machine translation (MT): instead of increasing parameter count or emitting explicit chain-of-thought tokens, they spend additional recurrent computation inside hidden states. We introduce LatentMT, the first systematic study of latent-reasoning LoopLMs for machine translation. LatentMT adapts a small 2.6B-parameter backbone model with lightweight training. Across 32 translation directions spanning high-, mid-, and low-resource languages, LatentMT achieves performance comparable to models three to five times larger. It is competitive in a high-resource language and achieves state-of-the-art performance on both mid-resource and low-resource languages. Studying the behavior of scaling the number of recurrent reasoning steps, we find that recurrent computation consistently improves translation quality in early steps, then saturates quickly afterwards. Our mechanistic analysis shows that hidden-representation differences shrink along the recurrent reasoning-step axis, supporting the observed saturation in performance. Finally, our efficiency analysis shows that LatentMT requires lower training and inference compute than much larger non-latent-reasoning models with similar performance, making latent recurrent computation a promising path toward compact, efficient, and strong machine translation.

Figures

Figures reproduced from arXiv: 2607.18618 by Chiyu Zhang, Muhammad Abdul-Mageed, Samar M. Magdy, Wei-Rui Chen, Wenhui Zhu, Zhipeng Wang.

Figure 1
Figure 1. Figure 1: Overview of LatentMT. A frozen LoopLM backbone is paired with compact per-language-pair LoRA adapters, avoiding full-model duplication. At inference, the shared backbone is unrolled for recur￾rent latent reasoning before each next-token prediction, adding internal computation without adding parameters. Swappable adapters let one deployed backbone serve many translation directions. Jeddi et al., 2026). This… view at source ↗
Figure 2
Figure 2. Figure 2: Performance scaling with the fixed number [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mechanistic saturation across recurrent tran [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Efficiency comparison of LatentMT against Qwen3-8B on DRT. Scaling computation through recurrent [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Human linguistic error analysis across recurrent depths for DRT and ArzEn-MultiGenre. Each panel [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Severity composition of tagged errors across recurrent depths, for DRT (left) and ArzEn-MultiGenre [PITH_FULL_IMAGE:figures/full_fig_p011_6.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

67 extracted references · 2 canonical work pages

  1. [1]

    2025 , eprint=

    Scaling Latent Reasoning via Looped Language Models , author=. 2025 , eprint=

    2025

  2. [2]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  3. [3]

    2025 , journal=

    Massively Multilingual Adaptation of Large Language Models Using Bilingual Translation Data , author=. 2025 , journal=

    2025

  4. [4]

    2025 , eprint=

    New Trends for Modern Machine Translation with Large Reasoning Models , author=. 2025 , eprint=

    2025

  5. [5]

    2026 , eprint=

    Chain-of-Thought Reasoning Improves Context-Aware Translation with Large Language Models , author=. 2026 , eprint=

    2026

  6. [7]

    2025 , eprint=

    Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach , author=. 2025 , eprint=

    2025

  7. [8]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , url =

    Wei, Jason and Wang, Xuezhi and Schuurmans, Dale and Bosma, Maarten and ichter, brian and Xia, Fei and Chi, Ed and Le, Quoc V and Zhou, Denny , booktitle =. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , url =

  8. [9]

    2025 , eprint=

    Training Large Language Models to Reason in a Continuous Latent Space , author=. 2025 , eprint=

    2025

  9. [10]

    2020 , eprint=

    Scaling Laws for Neural Language Models , author=. 2020 , eprint=

    2020

  10. [11]

    2026 , eprint=

    Unlocking Reasoning Capability on Machine Translation in Large Language Models , author=. 2026 , eprint=

    2026

  11. [12]

    2025 , eprint=

    Latent Chain-of-Thought? Decoding the Depth-Recurrent Transformer , author=. 2025 , eprint=

    2025

  12. [13]

    2023 , eprint=

    On Achieving Privacy-Preserving State-of-the-Art Edge Intelligence , author=. 2023 , eprint=

    2023

  13. [14]

    2023 , eprint=

    MobileNMT: Enabling Translation in 15MB and 30ms , author=. 2023 , eprint=

    2023

  14. [15]

    2020 , eprint=

    Edge Intelligence: Architectures, Challenges, and Applications , author=. 2020 , eprint=

    2020

  15. [16]

    Charlie Victor Snell and Jaehoon Lee and Kelvin Xu and Aviral Kumar , booktitle=. Scaling. 2025 , url=

    2025

  16. [17]

    2025 , eprint=

    SIM-CoT: Supervised Implicit Chain-of-Thought , author=. 2025 , eprint=

    2025

  17. [18]

    Edward J Hu and yelong shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle=. Lo. 2022 , url=

    2022

  18. [19]

    2025 , eprint=

    DeepTrans: Deep Reasoning Translation via Reinforcement Learning , author=. 2025 , eprint=

    2025

  19. [20]

    2025 , eprint=

    ExTrans: Multilingual Deep Reasoning Translation via Exemplar-Enhanced Reinforcement Learning , author=. 2025 , eprint=

    2025

  20. [21]

    2022 , eprint=

    No Language Left Behind: Scaling Human-Centered Machine Translation , author=. 2022 , eprint=

    2022

  21. [23]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    Small Data, Big Impact: Leveraging Minimal Data for Effective Machine Translation , author =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

  22. [24]

    COMET - Deploying a New State-of-the-art MT Evaluation Metric in Production

    Stewart, Craig and Rei, Ricardo and Farinha, Catarina and Lavie, Alon. COMET - Deploying a New State-of-the-art MT Evaluation Metric in Production. Proceedings of the 14th Conference of the Association for Machine Translation in the Americas (Volume 2: User Track). 2020

    2020

  23. [27]

    2025 , eprint=

    Compositional Translation: A Novel LLM-based Approach for Low-resource Machine Translation , author=. 2025 , eprint=

    2025

  24. [30]

    2026 , eprint=

    LLM Latent Reasoning as Chain of Superposition , author=. 2026 , eprint=

    2026

  25. [31]

    International Conference on Learning Representations , volume=

    Reasoning with latent thoughts: On the power of looped transformers , author=. International Conference on Learning Representations , volume=

  26. [32]

    Edge NLP for Efficient Machine Translation in Low Connectivity Areas , year=

    Watt, Tess and Chrysoulas, Christos and Gkatzia, Dimitra , booktitle=. Edge NLP for Efficient Machine Translation in Low Connectivity Areas , year=

  27. [33]

    Privacy Issues in Online Machine Translation Services - E uropean Perspective

    Kamocki, Pawel and O ' Regan, Jim. Privacy Issues in Online Machine Translation Services - E uropean Perspective. Proceedings of the Tenth International Conference on Language Resources and Evaluation ( LREC '16). 2016

    2016

  28. [34]

    31st International Conference on Distributed Computing Systems Workshops , pages=

    Finding a "Kneedle" in a Haystack: Detecting Knee Points in System Behavior , author=. 31st International Conference on Distributed Computing Systems Workshops , pages=

  29. [35]

    Forty-first International Conference on Machine Learning , year=

    Can Looped Transformers Learn to Implement Multi-step Gradient Descent for In-context Learning? , author=. Forty-first International Conference on Machine Learning , year=

  30. [36]

    The Fourteenth International Conference on Learning Representations , year=

    LoopFormer: Elastic-Depth Looped Transformers for Latent Reasoning via Shortcut Modulation , author=. The Fourteenth International Conference on Learning Representations , year=

  31. [38]

    Rania Al-Sabbagh. 2024. https://doi.org/10.1016/j.dib.2024.110271 Arzen-multigenre: An aligned parallel dataset of egyptian arabic song lyrics, novels, and subtitles, with english translations . Data in Brief, 54:110271

    arXiv 2024

  32. [39]

    Shabnam Ataee, Hugo Huart, and Andrei Popescu-Belis. 2026. https://arxiv.org/abs/2510.18077 Chain-of-thought reasoning improves context-aware translation with large language models . Preprint, arXiv:2510.18077

    arXiv 2026

  33. [40]

    Daphnee Chabal, Dolly Sapra, and Zoltán Ádám Mann. 2023. https://arxiv.org/abs/2302.05323 On achieving privacy-preserving state-of-the-art edge intelligence . Preprint, arXiv:2302.05323

    Pith/arXiv arXiv 2023

  34. [41]

    Jingcheng Deng, Liang Pang, Zihao Wei, Shicheng Xu, Zenghao Duan, Kun Xu, Yang Song, Huawei Shen, and Xueqi Cheng. 2026. https://arxiv.org/abs/2510.15522 Llm latent reasoning as chain of superposition . Preprint, arXiv:2510.15522

    arXiv 2026

  35. [42]

    Reddi, Stefanie Jegelka, and Sanjiv Kumar

    Khashayar Gatmiry, Nikunj Saunshi, Sashank J. Reddi, Stefanie Jegelka, and Sanjiv Kumar. 2024. https://openreview.net/forum?id=o8AaRKbP9K Can looped transformers learn to implement multi-step gradient descent for in-context learning? In Forty-first International Conference on Machine Learning

    2024

  36. [43]

    Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein

    Jonas Geiping, Sean McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R. Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, and Tom Goldstein. 2025. https://arxiv.org/abs/2502.05171 Scaling up test-time compute with latent reasoning: A recurrent depth approach . Preprint, arXiv:2502.05171

    Pith/arXiv arXiv 2025

  37. [44]

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2025. https://arxiv.org/abs/2412.06769 Training large language models to reason in a continuous latent space . Preprint, arXiv:2412.06769

    Pith/arXiv arXiv 2025

  38. [45]

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  39. [46]

    Ahmadreza Jeddi, Marco Ciccone, and Babak Taati. 2026. https://openreview.net/forum?id=RzYXb5YWBs Loopformer: Elastic-depth looped transformers for latent reasoning via shortcut modulation . In The Fourteenth International Conference on Learning Representations

    2026

  40. [47]

    Shaoxiong Ji, Zihao Li, Jaakko Paavola, Hengyu Luo, and Jörg Tiedemann. 2025. https://arxiv.org/abs/2506.00469 Massively multilingual adaptation of large language models using bilingual translation data . arXiv preprint 2506.00469

    arXiv 2025

  41. [48]

    Pawel Kamocki and Jim O ' Regan. 2016. https://aclanthology.org/L16-1706/ Privacy issues in online machine translation services - E uropean perspective . In Proceedings of the Tenth International Conference on Language Resources and Evaluation ( LREC '16) , pages 4458--4462, Portoro z , Slovenia. European Language Resources Association (ELRA)

    2016

  42. [49]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. https://arxiv.org/abs/2001.08361 Scaling laws for neural language models . Preprint, arXiv:2001.08361

    Pith/arXiv arXiv 2020

  43. [50]

    Ye Lin, Xiaohui Wang, Zhexi Zhang, Mingxuan Wang, Tong Xiao, and Jingbo Zhu. 2023. https://arxiv.org/abs/2306.04235 Mobilenmt: Enabling translation in 15mb and 30ms . Preprint, arXiv:2306.04235

    Pith/arXiv arXiv 2023

  44. [51]

    Sinuo Liu, Chenyang Lyu, Minghao Wu, Longyue Wang, Weihua Luo, Kaifu Zhang, and Zifu Shang. 2025. https://arxiv.org/abs/2503.10351 New trends for modern machine translation with large reasoning models . Preprint, arXiv:2503.10351

    Pith/arXiv arXiv 2025

  45. [52]

    Wenquan Lu, Yuechuan Yang, Kyle Lee, Yanshu Li, and Enqi Liu. 2025. https://arxiv.org/abs/2507.02199 Latent chain-of-thought? decoding the depth-recurrent transformer . Preprint, arXiv:2507.02199

    arXiv 2025

  46. [53]

    Magdy, Fakhraddin Alwajih, Abdellah El Mekki, Wesam El Sayed, and Muhammad Abdul-Mageed

    Samar M. Magdy, Fakhraddin Alwajih, Abdellah El Mekki, Wesam El Sayed, and Muhammad Abdul-Mageed. 2026. https://doi.org/10.18653/v1/2026.findings-acl.2012 LQM : Linguistically motivated multidimensional quality metrics for machine translation . In Findings of the A ssociation for C omputational L inguistics: ACL 2026 , pages 40470--40493, San Diego, Calif...

    work page doi:10.18653/v1/2026.findings-acl.2012 2026

  47. [54]

    Jean Maillard, Cynthia Gao, Elahe Kalbassi, Kaushik Ram Sadagopan, Vedanuj Goswami, Philipp Koehn, Angela Fan, and Francisco Guzmán. 2023. https://aclanthology.org/2023.acl-long.154 Small data, big impact: Leveraging minimal data for effective machine translation . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics ...

    2023

  48. [55]

    Lam Nguyen and Yang Xu. 2025. https://doi.org/10.18653/v1/2025.acl-srw.17 Reasoning for translation: Comparative analysis of chain-of-thought and tree-of-thought prompting for LLM translation . In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop), pages 259--275, Vienna, Austria. ...

    work page doi:10.18653/v1/2025.acl-srw.17 2025

  49. [56]

    Sara Rajaee, Sebastian Vincent, Alexandre Berard, Marzieh Fadaee, Kelly Marchisio, and Tom Kocmi. 2026. https://arxiv.org/abs/2602.14763 Unlocking reasoning capability on machine translation in large language models . Preprint, arXiv:2602.14763

    arXiv 2026

  50. [57]

    Ricardo Rei, Jos \'e G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and Andr \'e F. T. Martins. 2022 a . https://doi.org/10.18653/v1/2022.wmt-1.52 COMET -22: Unbabel- IST 2022 submission for the metrics shared task . In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 578...

    doi:10.18653/v1/2022.wmt-1.52 2022

  51. [58]

    Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, Jos \'e G

    Ricardo Rei, Marcos Treviso, Nuno M. Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, Jos \'e G. C. de Souza, Taisiya Glushkova, Duarte Alves, Luisa Coheur, Alon Lavie, and Andr \'e F. T. Martins. 2022 b . https://doi.org/10.18653/v1/2022.wmt-1.60 C omet K iwi: IST -unbabel 2022 submission for the quality estimation shared task . In Proceeding...

    doi:10.18653/v1/2022.wmt-1.60 2022

  52. [59]

    Ville Satopa, Jeannie Albrecht, David Irwin, and Barath Raghavan. 2011. Finding a "kneedle" in a haystack: Detecting knee points in system behavior. In 31st International Conference on Distributed Computing Systems Workshops, pages 166--171

    2011

  53. [60]

    Nikunj Saunshi, Nishanth Dikkala, Zhiyuan Li, Sanjiv Kumar, and Sashank J Reddi. 2025. Reasoning with latent thoughts: On the power of looped transformers. In International Conference on Learning Representations, volume 2025, pages 14855--14881

    2025

  54. [61]

    Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2025. https://openreview.net/forum?id=4FWAwZtd2n Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning . In The Thirteenth International Conference on Learning Representations

    2025

  55. [62]

    Craig Stewart, Ricardo Rei, Catarina Farinha, and Alon Lavie. 2020. https://aclanthology.org/2020.amta-user.4/ COMET - deploying a new state-of-the-art MT evaluation metric in production . In Proceedings of the 14th Conference of the Association for Machine Translation in the Americas (Volume 2: User Track), pages 78--109, Virtual. Association for Machine...

    2020

  56. [63]

    NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, and 20 others. 2022. https://arxiv.org/abs/2207.04672 No language...

    Pith/arXiv arXiv 2022

  57. [64]

    Jiaan Wang, Fandong Meng, Yunlong Liang, and Jie Zhou. 2025 a . https://doi.org/10.18653/v1/2025.findings-acl.351 DRT : Deep reasoning translation via long chain-of-thought . In Findings of the Association for Computational Linguistics: ACL 2025, pages 6770--6782, Vienna, Austria. Association for Computational Linguistics

    doi:10.18653/v1/2025.findings-acl.351 2025

  58. [65]

    Jiaan Wang, Fandong Meng, and Jie Zhou. 2025 b . https://arxiv.org/abs/2504.10187 Deeptrans: Deep reasoning translation via reinforcement learning . Preprint, arXiv:2504.10187

    Pith/arXiv arXiv 2025

  59. [66]

    Jiaan Wang, Fandong Meng, and Jie Zhou. 2025 c . https://arxiv.org/abs/2505.12996 Extrans: Multilingual deep reasoning translation via exemplar-enhanced reinforcement learning . Preprint, arXiv:2505.12996

    Pith/arXiv arXiv 2025

  60. [67]

    Tess Watt, Christos Chrysoulas, and Dimitra Gkatzia. 2023. https://doi.org/10.1109/WF-IoT58464.2023.10539577 Edge nlp for efficient machine translation in low connectivity areas . In 2023 IEEE 9th World Forum on Internet of Things (WF-IoT), pages 1--6

    arXiv 2023

  61. [68]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/9d5609613524ecf4f15af0f7b31abca4-Paper-Conference.pdf Chain-of-thought prompting elicits reasoning in large language models . In Advances in Neural Information Processing Systems...

    2022

  62. [69]

    Xilin Wei, Xiaoran Liu, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Jiaqi Wang, Xipeng Qiu, and Dahua Lin. 2025. https://arxiv.org/abs/2509.20317 Sim-cot: Supervised implicit chain-of-thought . Preprint, arXiv:2509.20317

    arXiv 2025

  63. [70]

    Di Wu, Seth Aycock, and Christof Monz. 2025. https://doi.org/10.18653/v1/2025.emnlp-main.1031 Please translate again: Two simple experiments on whether human-like reasoning helps translation . In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 20424--20440, Suzhou, China. Association for Computational Linguistics

    doi:10.18653/v1/2025.emnlp-main.1031 2025

  64. [71]

    Dianlei Xu, Tong Li, Yong Li, Xiang Su, Sasu Tarkoma, Tao Jiang, Jon Crowcroft, and Pan Hui. 2020. https://arxiv.org/abs/2003.12172 Edge intelligence: Architectures, challenges, and applications . Preprint, arXiv:2003.12172

    Pith/arXiv arXiv 2020

  65. [72]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. https://arxiv.org/abs/2505.09388 Qwen3 technical report . Preprint, arXiv:2505.09388

    Pith/arXiv arXiv 2025

  66. [73]

    Armel Zebaze, Benoît Sagot, and Rachel Bawden. 2025. https://arxiv.org/abs/2503.04554 Compositional translation: A novel llm-based approach for low-resource machine translation . Preprint, arXiv:2503.04554

    Pith/arXiv arXiv 2025

  67. [74]

    Rui-Jie Zhu, Zixuan Wang, Kai Hua, Tianyu Zhang, Ziniu Li, Haoran Que, Boyi Wei, Zixin Wen, Fan Yin, He Xing, Lu Li, Jiajun Shi, Kaijing Ma, Shanda Li, Taylor Kergan, Andrew Smith, Xingwei Qu, Mude Hui, Bohong Wu, and 14 others. 2025. https://arxiv.org/abs/2510.25741 Scaling latent reasoning via looped language models . Preprint, arXiv:2510.25741

    Pith/arXiv arXiv 2025