pith. machine review for the scientific record. sign in

arxiv: 2604.19144 · v1 · submitted 2026-04-21 · 💻 cs.CL

Recognition: unknown

ReflectMT: Internalizing Reflection for Efficient and High-Quality Machine Translation

Fandong Meng, Jinsong Su, Kunquan Li, Yingxue Zhang

Pith reviewed 2026-05-10 02:41 UTC · model grok-4.3

classification 💻 cs.CL
keywords machine translationreflection internalizationreinforcement learninglarge reasoning modelsefficient inferenceWMT24translate-reflect-refinedirect translation
0
0 comments X

The pith

A two-stage training process lets machine translation models internalize reflection to deliver better first-pass output than explicit reasoning models while using far fewer tokens.

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

The paper proposes ReflectMT as a two-stage reflection internalization method for machine translation that follows a translate-first-think-later paradigm instead of the usual think-first-then-translate approach used by large reasoning models. In the first stage reinforcement learning builds the model's ability to reflect on and refine translations, improving its semantic understanding. The second stage trains the model to absorb that reflection capability directly into its parameters so inference requires only a single direct translation pass. If correct, this would make high-quality machine translation practical at much lower cost and latency than current multi-step reasoning systems. Readers would care because explicit reasoning improves quality but creates prohibitive token usage and delays in real applications.

Core claim

ReflectMT uses a two-stage reflection internalization algorithm. The first stage employs reinforcement learning to develop the model's translate-reflect-refine capability and build semantic comprehension. The second stage internalizes the acquired knowledge so that at inference time the model produces high-quality translations in a single direct pass without any explicit reasoning steps. On WMT24 this yields first-pass outputs that outperform multi-step reasoning models such as DeepSeek-R1 in both automatic metrics and GPT-based evaluation, with a 2.16-point GPT quality gain and 94.33 percent reduction in token consumption.

What carries the argument

The two-stage reflection internalization algorithm, which first cultivates explicit translate-reflect-refine capability via reinforcement learning and then folds that capability into the model's direct translation parameters.

If this is right

  • Inference operates in direct translation mode with no explicit reasoning steps yet still produces higher quality than multi-step reasoning models.
  • Token consumption drops by 94.33 percent compared with explicit reasoning approaches while improving GPT-evaluated quality by 2.16 points on WMT24.
  • The model gains enhanced semantic comprehension and task-specific knowledge from the initial reflection training stage.
  • Latency and computational overhead from visible reasoning trajectories are eliminated during deployment.

Where Pith is reading between the lines

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

  • The same internalization pattern could be applied to other reasoning-intensive tasks such as code generation or mathematical problem solving to cut inference costs.
  • Production machine translation pipelines would see major reductions in latency and resource requirements if the direct-pass quality holds across languages.
  • Further tests on out-of-domain or low-resource language pairs could show whether internalized reflection generalizes better than explicit step-by-step methods.

Load-bearing premise

The second stage successfully transfers quality gains from explicit reflection training into the model's direct translation behavior without loss or introduction of new biases.

What would settle it

A head-to-head test on WMT24 where ReflectMT's single-pass outputs fail to show at least a 2-point GPT quality improvement and at least 90 percent token reduction relative to DeepSeek-R1's multi-step outputs would falsify the performance claim.

Figures

Figures reproduced from arXiv: 2604.19144 by Fandong Meng, Jinsong Su, Kunquan Li, Yingxue Zhang.

Figure 1
Figure 1. Figure 1: Training vs. Inference Paradigm of ReflectMT. During training, the model generates the complete [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Reflection Internalization Framework. (a) Data Construction: A multi-agent system [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training Dynamics: As the training steps increase (a) indicates the number of refinements; (b) represents [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
read the original abstract

Recent years have witnessed growing interest in applying Large Reasoning Models (LRMs) to Machine Translation (MT). Existing approaches predominantly adopt a "think-first-then-translate" paradigm. Although explicit reasoning trajectories significantly enhance translation quality, they incur prohibitive inference costs and latency. To address these limitations, we propose ReflectMT, a two-stage reflection internalization algorithm for machine translation that employs a "translate-first-think-later" paradigm. Our approach develops the model's "translate-reflect-refine" capability through reinforcement learning. In the first stage, we cultivate the model's capacity for high-quality reflection and refinement, thereby enhancing its semantic comprehension and task-specific knowledge. In the second stage, we train the model to internalize the knowledge acquired during reflection. As a result, during inference, ReflectMT operates in a direct translation mode, producing high-quality translations on the first attempt without any explicit reasoning steps. Experimental results on datasets such as WMT24 demonstrate that our model's first-pass translations during inference outperform multi-step reasoning LRMs such as DeepSeek-R1 in both automatic metrics and GPT-based evaluation, achieving a 2.16-point improvement in GPT-based translation quality evaluation while reducing token consumption by 94.33%.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes ReflectMT, a two-stage reflection internalization algorithm for machine translation. Stage 1 employs reinforcement learning to cultivate a 'translate-reflect-refine' capability that enhances semantic comprehension. Stage 2 internalizes the acquired knowledge so that inference reduces to direct translation without explicit reasoning steps. On benchmarks such as WMT24, the resulting first-pass outputs are claimed to surpass multi-step reasoning LRMs (e.g., DeepSeek-R1) in both automatic metrics and GPT-based quality evaluation (by 2.16 points) while cutting token consumption by 94.33%.

Significance. If the internalization step successfully compresses reflection gains into direct-generation parameters without introducing new biases or quality loss, the work would offer a practical route to high-quality MT at the inference cost of standard direct models. This addresses a central limitation of current LRM-based translation approaches and could influence how reasoning behaviors are distilled in production systems. The two-stage 'translate-first-think-later' framing is a clear conceptual contribution, though its empirical support remains to be demonstrated through controlled experiments.

major comments (2)
  1. [Abstract] Abstract: The central claim that post-internalization first-pass translations outperform DeepSeek-R1 by 2.16 GPT points and 94.33% token reduction rests on the unverified assumption that stage-2 internalization preserves stage-1 reflection quality. No ablation is reported that compares direct-translation performance after internalization against the explicit multi-step outputs from the same base model trained only on stage 1.
  2. [Abstract] Abstract: The reported benchmark gains are presented without any description of the base model, RL algorithm, training data, hyper-parameters, or statistical significance tests. This absence prevents verification that the improvements are attributable to the proposed internalization rather than confounding factors such as data scale or model capacity.
minor comments (1)
  1. [Abstract] Abstract: The phrase 'datasets such as WMT24' should be replaced by an explicit enumeration of all evaluation sets together with their language pairs and sizes to allow readers to assess coverage.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. The comments correctly identify areas where additional empirical controls and clarity would strengthen the claims regarding reflection internalization. We address each major comment below and will revise the manuscript to incorporate the suggested improvements.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that post-internalization first-pass translations outperform DeepSeek-R1 by 2.16 GPT points and 94.33% token reduction rests on the unverified assumption that stage-2 internalization preserves stage-1 reflection quality. No ablation is reported that compares direct-translation performance after internalization against the explicit multi-step outputs from the same base model trained only on stage 1.

    Authors: We agree that a controlled ablation comparing the internalized direct-translation model (stage 2) against the explicit multi-step reasoning outputs from the identical base model after only stage 1 would provide direct evidence that internalization preserves the quality gains. Our current evaluation demonstrates that ReflectMT's first-pass translations surpass the external baseline DeepSeek-R1 in both quality and efficiency. To rigorously test the preservation assumption, we will add this ablation to the revised manuscript, reporting results on the same base model, training regime, and evaluation metrics for both the stage-1 explicit-reasoning variant and the full two-stage internalized model. revision: yes

  2. Referee: [Abstract] Abstract: The reported benchmark gains are presented without any description of the base model, RL algorithm, training data, hyper-parameters, or statistical significance tests. This absence prevents verification that the improvements are attributable to the proposed internalization rather than confounding factors such as data scale or model capacity.

    Authors: The full manuscript contains the experimental details on the base model, the RL algorithms used in each stage, the training data construction, hyper-parameter choices, and statistical significance testing. To address the concern about self-contained presentation of the benchmark claims, we will revise the abstract to include a brief description of the base model and RL setup. We will also ensure the main text explicitly links the observed gains to the internalization process through the controlled comparisons already present in the experiments. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical claims rest on external benchmarks

full rationale

The paper describes a two-stage RL-based training procedure (reflection cultivation followed by internalization) whose outputs are evaluated on external WMT24 data using automatic metrics and GPT-based human-like scoring against independent baselines such as DeepSeek-R1. No equations, predictions, or first-principles results are presented that reduce by construction to the method's own inputs or fitted quantities. The reported gains (2.16 GPT points, 94.33% token reduction) are measured quantities, not quantities defined by the training stages themselves. No self-citation chains, uniqueness theorems, or ansatzes are invoked to justify core claims.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that reinforcement learning can successfully internalize multi-step reflection behavior into a single forward pass.

axioms (1)
  • domain assumption Reinforcement learning can train models to internalize complex behaviors such as reflection and refinement.
    This underpins both training stages described in the abstract.

pith-pipeline@v0.9.0 · 5515 in / 1139 out tokens · 42467 ms · 2026-05-10T02:41:33.451639+00:00 · methodology

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

76 extracted references · 45 canonical work pages · 15 internal anchors

  1. [1]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    2024 , howpublished =

    OpenAI , title =. 2024 , howpublished =

    2024

  3. [3]

    2023 , eprint=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. 2023 , eprint=

    2023

  4. [4]

    2023 , eprint=

    Towards Mitigating Hallucination in Large Language Models via Self-Reflection , author=. 2023 , eprint=

    2023

  5. [5]

    2023 , eprint=

    Large Language Models are Better Reasoners with Self-Verification , author=. 2023 , eprint=

    2023

  6. [6]

    2024 , eprint=

    TasTe: Teaching Large Language Models to Translate through Self-Reflection , author=. 2024 , eprint=

    2024

  7. [7]

    2025 , eprint=

    Instruct-of-Reflection: Enhancing Large Language Models Iterative Reflection Capabilities via Dynamic-Meta Instruction , author=. 2025 , eprint=

    2025

  8. [8]

    2023 , eprint=

    Scaling up COMETKIWI: Unbabel-IST 2023 Submission for the Quality Estimation Shared Task , author=. 2023 , eprint=

    2023

  9. [9]

    2025 , eprint=

    DeepSeek-V3 Technical Report , author=. 2025 , eprint=

    2025

  10. [10]

    Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1) , pages=

    Results of the WMT19 metrics shared task: Segment-level and strong MT systems pose big challenges , author=. Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1) , pages=

  11. [11]

    2021 , eprint=

    LoRA: Low-Rank Adaptation of Large Language Models , author=. 2021 , eprint=

    2021

  12. [12]

    2024 , eprint=

    MetricX-24: The Google Submission to the WMT 2024 Metrics Shared Task , author=. 2024 , eprint=

    2024

  13. [13]

    2022 , eprint=

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

    2022

  14. [14]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  15. [15]

    2026 , eprint=

    Enhancing Self-Correction in Large Language Models through Multi-Perspective Reflection , author=. 2026 , eprint=

    2026

  16. [16]

    arXiv preprint arXiv:2505.12996 , year=

    Extrans: Multilingual deep reasoning translation via exemplar-enhanced reinforcement learning , author=. arXiv preprint arXiv:2505.12996 , year=

    work page arXiv

  17. [17]

    2025 , eprint=

    SlangDIT: Benchmarking LLMs in Interpretative Slang Translation , author=. 2025 , eprint=

    2025

  18. [18]

    Gokul Swamy, Sanjiban Choudhury, Wen Sun, Zhiwei Steven Wu, and J

    All roads lead to likelihood: The value of reinforcement learning in fine-tuning , author=. arXiv preprint arXiv:2503.01067 , year=

    work page arXiv

  19. [19]

    o3-mini: How Well can Reasoning LLMs Evaluate MT and Summarization? , author=

    DeepSeek vs. o3-mini: How Well can Reasoning LLMs Evaluate MT and Summarization? , author=. arXiv preprint arXiv:2504.08120 , year=

    work page arXiv

  20. [20]

    Qwen3 Technical Report

    Qwen3 Technical Report , author=. arXiv preprint arXiv:2505.09388 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    , author=

    Measuring nominal scale agreement among many raters. , author=. Psychological bulletin , volume=. 1971 , publisher=

    1971

  22. [22]

    Best-Worst Scaling More Reliable than Rating Scales: A Case Study on Sentiment Intensity Annotation

    Kiritchenko, Svetlana and Mohammad, Saif. Best-Worst Scaling More Reliable than Rating Scales: A Case Study on Sentiment Intensity Annotation. Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2017. doi:10.18653/v1/P17-2074

    work page doi:10.18653/v1/p17-2074 2017

  23. [23]

    2024 , journal =

    HybridFlow: A Flexible and Efficient RLHF Framework , author =. 2024 , journal =

    2024

  24. [24]

    arXiv preprint arXiv:2402.13116 , year =

    A survey on knowledge distillation of large language models , author=. arXiv preprint arXiv:2402.13116 , year=

    work page arXiv

  25. [25]

    Literary Machine Translation under the Magnifying Glass: Assessing the Quality of an NMT -Translated Detective Novel on Document Level

    Fonteyne, Margot and Tezcan, Arda and Macken, Lieve. Literary Machine Translation under the Magnifying Glass: Assessing the Quality of an NMT -Translated Detective Novel on Document Level. Proceedings of the Twelfth Language Resources and Evaluation Conference. 2020

    2020

  26. [26]

    Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models

    Towards reasoning era: A survey of long chain-of-thought for reasoning large language models , author=. arXiv preprint arXiv:2503.09567 , year=

    work page internal anchor Pith review arXiv

  27. [27]

    From System 1 to System 2: A Survey of Reasoning Large Language Models

    From system 1 to system 2: A survey of reasoning large language models , author=. arXiv preprint arXiv:2502.17419 , year=

    work page internal anchor Pith review arXiv

  28. [28]

    o1-coder: an o1 replication for coding

    o1-coder: an o1 replication for coding , author=. arXiv preprint arXiv:2412.00154 , year=

    work page arXiv

  29. [29]

    arXiv preprint arXiv:2502.01142 (2025),https://arxiv.org/abs/2502.01142

    DeepRAG: Thinking to Retrieval Step by Step for Large Language Models , author=. arXiv preprint arXiv:2502.01142 , year=

    work page arXiv

  30. [30]

    arXiv e-prints , pages=

    Drt-o1: Optimized deep reasoning translation via long chain-of-thought , author=. arXiv e-prints , pages=

  31. [31]

    Marco-o1: Towards open reasoning models for open-ended solutions

    Marco-o1: Towards open reasoning models for open-ended solutions , author=. arXiv preprint arXiv:2411.14405 , year=

    work page arXiv

  32. [32]

    Evaluating o1-like llms: Unlocking reasoning for translation through comprehensive analysis

    Evaluating o1-like llms: Unlocking reasoning for translation through comprehensive analysis , author=. arXiv preprint arXiv:2502.11544 , year=

    work page arXiv

  33. [33]

    arXiv preprint arXiv:2502.19735 , year=

    R1-T1: Fully Incentivizing Translation Capability in LLMs via Reasoning Learning , author=. arXiv preprint arXiv:2502.19735 , year=

    work page arXiv

  34. [34]

    COMET : A Neural Framework for MT Evaluation

    Rei, Ricardo and Stewart, Craig and Farinha, Ana C and Lavie, Alon. COMET : A Neural Framework for MT Evaluation. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020. doi:10.18653/v1/2020.emnlp-main.213

    work page doi:10.18653/v1/2020.emnlp-main.213 2020

  35. [35]

    REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalization

    REINFORCE++: A Simple and Efficient Approach for Aligning Large Language Models , author=. arXiv preprint arXiv:2501.03262 , year=

    work page internal anchor Pith review arXiv

  36. [36]

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale , author=. arXiv preprint arXiv:2503.14476 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2503.10351 , year=

    New Trends for Modern Machine Translation with Large Reasoning Models , author=. arXiv preprint arXiv:2503.10351 , year=

    work page arXiv

  38. [38]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  39. [39]

    Large Language Models Effectively Leverage Document-level Context for Literary Translation, but Critical Errors Persist

    Karpinska, Marzena and Iyyer, Mohit. Large Language Models Effectively Leverage Document-level Context for Literary Translation, but Critical Errors Persist. Proceedings of the Eighth Conference on Machine Translation. 2023. doi:10.18653/v1/2023.wmt-1.41

    work page doi:10.18653/v1/2023.wmt-1.41 2023

  40. [40]

    Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning

    Search-r1: Training llms to reason and leverage search engines with reinforcement learning , author=. arXiv preprint arXiv:2503.09516 , year=

    work page Pith review arXiv

  41. [41]

    B leu: a Method for Automatic Evaluation of Machine Translation

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing. B leu: a Method for Automatic Evaluation of Machine Translation. Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. 2002. doi:10.3115/1073083.1073135

    work page doi:10.3115/1073083.1073135 2002

  42. [42]

    WebGPT: Browser-assisted question-answering with human feedback

    Webgpt: Browser-assisted question-answering with human feedback , author=. arXiv preprint arXiv:2112.09332 , year=

    work page internal anchor Pith review arXiv

  43. [43]

    Advances in neural information processing systems , volume=

    Learning to summarize with human feedback , author=. Advances in neural information processing systems , volume=

  44. [44]

    and Zerva, Chrysoula and Farinha, Ana C and Maroti, Christine and C

    Rei, Ricardo and Treviso, Marcos and Guerreiro, Nuno M. and Zerva, Chrysoula and Farinha, Ana C and Maroti, Christine and C. de Souza, Jos \'e G. and Glushkova, Taisiya and Alves, Duarte and Coheur, Luisa and Lavie, Alon and Martins, Andr \'e F. T. C omet K iwi: IST -Unbabel 2022 Submission for the Quality Estimation Shared Task. Proceedings of the Sevent...

    2022

  45. [45]

    BLEURT : Learning Robust Metrics for Text Generation

    Sellam, Thibault and Das, Dipanjan and Parikh, Ankur. BLEURT : Learning Robust Metrics for Text Generation. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020. doi:10.18653/v1/2020.acl-main.704

    work page doi:10.18653/v1/2020.acl-main.704 2020

  46. [46]

    arXiv preprint arXiv:2411.16594 , year=

    From generation to judgment: Opportunities and challenges of llm-as-a-judge , author=. arXiv preprint arXiv:2411.16594 , year=

    work page arXiv

  47. [47]

    MM-Eval : A multilingual meta-evaluation benchmark for LLM -as-a-judge and reward models

    MM-Eval: A Multilingual Meta-Evaluation Benchmark for LLM-as-a-Judge and Reward Models , author=. arXiv preprint arXiv:2410.17578 , year=

    work page arXiv

  48. [48]

    Multilingual Machine Translation with Large Language Models: Empirical Results and Analysis

    Zhu, Wenhao and Liu, Hongyi and Dong, Qingxiu and Xu, Jingjing and Huang, Shujian and Kong, Lingpeng and Chen, Jiajun and Li, Lei. Multilingual Machine Translation with Large Language Models: Empirical Results and Analysis. Findings of the Association for Computational Linguistics: NAACL 2024. 2024. doi:10.18653/v1/2024.findings-naacl.176

    work page doi:10.18653/v1/2024.findings-naacl.176 2024

  49. [49]

    arXiv preprint arXiv:2504.10187 , year=

    Deep Reasoning Translation via Reinforcement Learning , author=. arXiv preprint arXiv:2504.10187 , year=

    work page arXiv

  50. [50]

    arXiv preprint arXiv:2504.10160 , year=

    MT-R1-Zero: Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning , author=. arXiv preprint arXiv:2504.10160 , year=

    work page arXiv

  51. [51]

    Search-o1: Agentic Search-Enhanced Large Reasoning Models

    Search-o1: Agentic search-enhanced large reasoning models , author=. arXiv preprint arXiv:2501.05366 , year=

    work page internal anchor Pith review arXiv

  52. [52]

    Is C hat GPT a Good NLG Evaluator? A Preliminary Study

    Wang, Jiaan and Liang, Yunlong and Meng, Fandong and Sun, Zengkui and Shi, Haoxiang and Li, Zhixu and Xu, Jinan and Qu, Jianfeng and Zhou, Jie. Is C hat GPT a Good NLG Evaluator? A Preliminary Study. Proceedings of the 4th New Frontiers in Summarization Workshop. 2023. doi:10.18653/v1/2023.newsum-1.1

    work page doi:10.18653/v1/2023.newsum-1.1 2023

  53. [53]

    DeepSeek-V3 Technical Report

    Deepseek-v3 technical report , author=. arXiv preprint arXiv:2412.19437 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  54. [54]

    Large Language Models Are State-of-the-Art Evaluators of Translation Quality

    Kocmi, Tom and Federmann, Christian. Large Language Models Are State-of-the-Art Evaluators of Translation Quality. Proceedings of the 24th Annual Conference of the European Association for Machine Translation. 2023

    2023

  55. [55]

    Qwen2.5 Technical Report

    Qwen2. 5 technical report , author=. arXiv preprint arXiv:2412.15115 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  56. [56]

    Proximal Policy Optimization Algorithms

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  57. [57]

    The Llama 3 Herd of Models

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  58. [58]

    Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

    Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters , author=. Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

  59. [59]

    doi: 10.18653/v1/2024.acl-demos.38

    Zheng, Yaowei and Zhang, Richong and Zhang, Junhao and Ye, Yanhan and Luo, Zheyan. L lama F actory: Unified Efficient Fine-Tuning of 100+ Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). 2024. doi:10.18653/v1/2024.acl-demos.38

    work page doi:10.18653/v1/2024.acl-demos.38 2024

  60. [60]

    Hugging Face , year=

    Qwq: Reflect deeply on the boundaries of the unknown , author=. Hugging Face , year=

  61. [61]

    GPT-4o System Card

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  62. [62]

    International Conference on Learning Representations , year=

    Sequence level training with recurrent neural networks , author=. International Conference on Learning Representations , year=

  63. [63]

    International Conference on Learning Representations , year=

    An Actor-Critic Algorithm for Sequence Prediction , author=. International Conference on Learning Representations , year=

  64. [64]

    Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation

    Google's neural machine translation system: Bridging the gap between human and machine translation , author=. arXiv preprint arXiv:1609.08144 , year=

    work page internal anchor Pith review arXiv

  65. [65]

    A Study of Reinforcement Learning for Neural Machine Translation

    Wu, Lijun and Tian, Fei and Qin, Tao and Lai, Jianhuang and Liu, Tie-Yan. A Study of Reinforcement Learning for Neural Machine Translation. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018. doi:10.18653/v1/D18-1397

    work page doi:10.18653/v1/d18-1397 2018

  66. [66]

    Minimum Risk Training for Neural Machine Translation

    Shen, Shiqi and Cheng, Yong and He, Zhongjun and He, Wei and Wu, Hua and Sun, Maosong and Liu, Yang. Minimum Risk Training for Neural Machine Translation. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2016. doi:10.18653/v1/P16-1159

    work page doi:10.18653/v1/p16-1159 2016

  67. [67]

    Exploring Document-Level Literary Machine Translation with Parallel Paragraphs from World Literature

    Thai, Katherine and Karpinska, Marzena and Krishna, Kalpesh and Ray, Bill and Inghilleri, Moira and Wieting, John and Iyyer, Mohit. Exploring Document-Level Literary Machine Translation with Parallel Paragraphs from World Literature. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. doi:10.18653/v1/2022.emnlp-main.672

    work page doi:10.18653/v1/2022.emnlp-main.672 2022

  68. [68]

    Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing , pages=

    “Poetic” statistical machine translation: rhyme and meter , author=. Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing , pages=

    2010

  69. [69]

    Proceedings of the 7th Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities , pages=

    The (un) faithful machine translator , author=. Proceedings of the 7th Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities , pages=

  70. [70]

    Fourth Workshop on Computational Linguistics for Literature-co-located with NAACL 2015 , year=

    Automated translation of a literary work: a pilot study , author=. Fourth Workshop on Computational Linguistics for Literature-co-located with NAACL 2015 , year=

    2015

  71. [71]

    Frontiers in Digital Humanities , volume=

    Post-editing effort of a novel with statistical and neural machine translation , author=. Frontiers in Digital Humanities , volume=. 2018 , publisher=

    2018

  72. [72]

    On the Weaknesses of Reinforcement Learning for Neural Machine Translation , author=

  73. [73]

    Revisiting the Weaknesses of Reinforcement Learning for Neural Machine Translation , url=

    Kiegeland, Samuel and Kreutzer, Julia. Revisiting the Weaknesses of Reinforcement Learning for Neural Machine Translation. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021. doi:10.18653/v1/2021.naacl-main.133

    work page doi:10.18653/v1/2021.naacl-main.133 2021

  74. [74]

    Dynamic Context Selection for Document-level Neural Machine Translation via Reinforcement Learning

    Kang, Xiaomian and Zhao, Yang and Zhang, Jiajun and Zong, Chengqing. Dynamic Context Selection for Document-level Neural Machine Translation via Reinforcement Learning. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020. doi:10.18653/v1/2020.emnlp-main.175

    work page doi:10.18653/v1/2020.emnlp-main.175 2020

  75. [75]

    arXiv preprint arXiv:2602.18884 , year=

    TPRU: Advancing Temporal and Procedural Understanding in Large Multimodal Models , author=. arXiv preprint arXiv:2602.18884 , year=

    work page arXiv

  76. [76]

    arXiv preprint arXiv:2602.22227 , year=

    To Deceive is to Teach? Forging Perceptual Robustness via Adversarial Reinforcement Learning , author=. arXiv preprint arXiv:2602.22227 , year=

    work page arXiv