Pith. sign in

REVIEW 4 major objections 6 minor 44 references

M2PO enables a 7B translation model to rival GPT-4o by adding a word-alignment faithfulness penalty to quality-estimation rewards and training on multiple preference pairs from the same candidate pool.

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 ↗

M2PO combines a QE-plus-alignment reward with a dynamic curriculum and multi-pair DPO loss, and reports WMT21-22 gains for a 7B translation model, but the abstract's WMT23/24 9B parity claims are unsupported.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Solid engineering with a real idea, but the abstract oversells and the faithfulness signal is not validated. the 4 major comments →

arxiv 2510.13434 v2 pith:55ILPIW2 submitted 2025-10-15 cs.CL

$M^2PO$: Multi-Perspective Multi-Pair Preference Optimization for Machine Translation

Hao Wang , Linlong Xu , Heng Liu , Yangyang Liu , Xiaohu Zhao , Bo Zeng , Liangying Shao , Yichen Dong
show 6 more authors
Xinwei Wu Jiang Zhou Tianyu Dong xiangxiang Zeng Longyue Wang Weihua Luo
This is my paper
classification cs.CL
keywords machine translationpreference optimizationdirect preference optimizationhallucinationfaithfulnessquality estimationmulti-pair learningLLM alignment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper argues that current preference-optimization methods for machine translation fail because quality-estimation (QE) scores ignore partial hallucinations and omissions, favoring fluent but unfaithful translations. M2PO corrects this with a composite reward that adds a word-alignment-based factuality score to the QE score, a curriculum that gradually shifts from static expert scores to the model's own judgment, and a multi-pair objective that uses every candidate instead of a single best-worst pair. On WMT21-22, the 7B M2PO model outperforms all open-source baselines, matches GPT-4o, and surpasses its own data source GPT-4o-mini. If the results hold, the framework offers a data-efficient, label-free way to align small open models to high-fidelity translation.

Core claim

The central claim is that QE-based rewards are systematically blind to partial errors — partial hallucinations and omissions — and that this blind spot can be removed by supplementing the QE score with a word-alignment faithfulness score (S_align from WSPAlign). The paper then shows that with this corrected reward, using the full ranked candidate list in K/2 head-to-tail pairs, weighted by the reward gap, yields better alignment than the standard single best-vs-worst comparison. The reported outcome is a 7B ALMA-based model that advances simultaneously on translation quality (XCOMET) and faithfulness (coverage score), becoming competitive with GPT-4o and beating GPT-4o-mini on WMT21-22. The

What carries the argument

The paper's load-bearing mechanism is the composite static reward r_s = r_qe + λ_f·S_align, where r_qe is a QE score (KIWI-XXL) and S_align is a word-level alignment score (WSPAlign) that penalizes content not grounded in the source. During training, z-score normalized static scores are fused with the model's own log-probabilities via r_fused = (1-α_t)·z(r_s) + α_t·z(log π_θ), with α_t increasing from 0.1 to 0.9, so the model gradually trusts its own judgment. Candidates are ranked by r_fused and split into K/2 head-to-tail pairs; the training loss combines (1) a weighted multi-pair DPO loss with softmax weights on reward differences, (2) a ListNet-style listwise ranking loss that keeps the

Load-bearing premise

The load-bearing premise is that the automatic faithfulness scores (WSPAlign's alignment score during training and Gemini's coverage score during evaluation) genuinely reflect human judgments of partial hallucinations and omissions; without human evaluation, the reported faithfulness gains could be an artifact of using similar automatic signals on both sides.

What would settle it

A human evaluation on a benchmark designed for partial errors, such as HalOmi, comparing M2PO versus DPO/CPO translations on the same source sentences would settle the claim: if human raters find no faithfulness superiority for M2PO, or find it less faithful, the central claim fails. A cheaper check is to rerun the preference construction with an independent human-validated faithfulness metric in place of WSPAlign and see whether the criterion still selects the same winners.

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

If this is right

  • A corrected reward signal can be reused: the same composite reward (QE + alignment bonus) can be plugged into any DPO-like algorithm, and the paper reports consistent gains across DPO, KTO, SimPO, ORPO, and CPO.
  • Small models can beat their teachers: by learning from a richer preference signal rather than imitating outputs, a 7B model surpasses GPT-4o-mini, one of the models that generated its training candidates.
  • Fluency collapse is the main risk in multi-pair alignment: ablating the behavior-cloning term costs more than five XCOMET points, indicating that a simple NLL safeguard on the best candidate is essential for multi-pair DPO-style training.
  • Faithfulness and fluency are not in inherent conflict in preference optimization; the M2PO model rises on both XCOMET and the coverage score simultaneously, decoupling the two axes.
  • The QE blind-spot finding means that any preference dataset built with QE scores alone will contain systematically mislabeled winner-loser pairs for partial hallucinations, which may drive the 'fluency bias' observed in many aligned MT models.

Where Pith is reading between the lines

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

  • The paper evaluates faithfulness with an LLM-judge coverage score from Gemini-2.0-Flash, but no human evaluation is reported; the strongest test would be a human study on partial hallucinations to confirm that the automatic faithfulness gains reflect real fidelity improvements rather than agreement between two correlated automatic metrics.
  • The curriculum that shifts from external static scores to the model's own log-probabilities is reminiscent of an exploration-exploitation trade-off; it could generalize to any alignment pipeline where the reward model is known to be biased, but the paper does not test it outside MT.
  • The head-to-tail pairing strategy is one of several possible ways to exploit a ranked candidate list; an immediate extension would be to vary K and the pairing scheme and see where gains saturate, or to generate the entire candidate pool from a single model at multiple temperatures rather than mixing in GPT-4o-mini.
  • If the QE blind spot is as systematic as the paper's HalOmi analysis suggests, the same multi-perspective reward design could be applied to other faithfulness-critical generation tasks such as summarization and instruction following, where partial omission and hallucination are also under-penalized by reference-free quality metrics.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes M2PO, a multi-pair, multi-perspective preference optimization framework for machine translation. It constructs candidate pools from an SFT model and GPT-4o-mini, augments them with difficult negatives, and scores candidates with a static reward r_s = r_qe + λ_f S_align (Eq. 2) that combines a QE score (KIWI-XXL) with a WSPAlign-based factuality bonus. During training, the static score is fused with the model's own log-probabilities via a curriculum weight α_t (Eq. 3), and the ranked candidates are used to create K/2 head-to-tail preference pairs. The final loss (Eq. 4) combines a multi-pair DPO-style loss, a listwise ranking loss, and a behavior-cloning loss. Experiments on WMT21-22 with ALMA-7B-LoRA report improvements over DPO/CPO baselines and competitiveness with GPT-4o and GPT-4o-mini. The arXiv abstract additionally claims results on WMT23, WMT24, FLORES-200, and a 9B model with parity to Gemini-2.0-Flash, but these are not present in the body.

Significance. The core idea is plausible and potentially useful: constructing a diverse candidate pool with difficult negatives, combining a QE score with an alignment-based factuality bonus, fusing static and self-reward scores with a curriculum, and training on multiple pairs are all sensible design choices. Table 1 provides a useful demonstration that standard QE metrics lose correlation with human labels when partial hallucination examples are included, and the authors are careful not to reuse the training-time metric (KIWI-XXL) for evaluation. If the claims are reproducible, M2PO would be a valuable data-centric recipe for MT alignment. However, the manuscript does not validate the faithfulness signal against human judgments, lacks statistical significance testing, contains a formula error in its core loss, and includes a material mismatch between the abstract and the reported experiments. The stated plan to release code and data is a strength, but availability is not yet confirmed.

major comments (4)
  1. [§4.4, Eq. (5)] Equation (5) defines L_DM-DPO = -Σ_i w_i log σ(β Δ_i log πθ). This contains only the policy-model log-ratio; the standard DPO objective in Eq. (1) includes the reference-model term Δπ_ref = log π_ref(y_w|x) - log π_ref(y_l|x). As written, Eq. (5) is a reference-free/SimPO-like loss, not DPO. The method therefore cannot be reproduced from the equations, and the statement that M2PO wraps 'DPO-like algorithms' is misleading. Please either insert the π_ref term, or explicitly define the loss as reference-free and justify the choice. If the implemented loss differs from Eq. (5), all main experiments and ablations need to be re-examined.
  2. [§4.3, §5.2, Appendix A] The faithfulness component of the preference signal is S_align from WSPAlign in Eq. (2), yet §3 validates only the Gemini Coverage score against HalOmi; WSPAlign is never evaluated against human annotations. In addition, the final faithfulness evaluation (§5.2, Appendix A) uses a Gemini-2.0-Flash coverage score from the same family as the motivating analysis. No human evaluation of the final translations is reported anywhere. If S_align, or the Gemini judge, does not track human judgments of partial hallucinations/omissions, the preference pairs are mislabeled and the reported faithfulness gains may be an artifact of optimizing a circular reward. The authors should validate S_align on HalOmi and include a human faithfulness evaluation of the M2PO outputs.
  3. [§5.2, Tables 2-3] Tables 2-3 and Table 7 report single-run point estimates with no error bars or significance tests. Several headline differences are very small (e.g., Table 2 En→De XCOMET: M2PO 97.25 vs CPO 97.11; En→Cs COMET22: 90.40 vs 90.29), which is within typical COMET/XCOMET variability. The repeated claim that M2PO 'significantly' or 'substantially' outperforms baselines is therefore unsupported. Please provide multiple seeds and pairwise significance tests, or weaken the claims accordingly.
  4. [Abstract vs §5.1, §5.3] The arXiv abstract promises experiments on WMT23, WMT24 and FLORES-200 with a 9B model that reaches parity with GPT-4o and Gemini-2.0-Flash. The full text's own abstract and §5 only report WMT21-22, ALMA-7B-LoRA, and baselines GPT-4o/GPT-4o-mini; FLORES-200 is used only to build the preference dataset (§5.1), and Gemini-2.0-Flash appears only as the name of the coverage-score judge. This is a material mismatch between the claimed scope and the evidence. Either add the promised experiments or correct the abstract to match the body; as written, the headline claim cannot be verified.
minor comments (6)
  1. [§4.4] The index i in Eq. (5) and the notation Δ_i are not defined; specify i = 1,...,K/2 and define Δ_i log πθ = log πθ(y_w^i) - log πθ(y_l^i). Also specify how λ_pref, λ_rank, and λ_bc are chosen.
  2. [§5.1] Hyperparameters λ_f, β, α_t schedule, τ_s, τ_w, and K are set ad hoc; provide at least a sensitivity analysis for λ_f and β, since the preference-pair construction depends on them.
  3. [Figure 1] The 'volatility' claim is based on visual inspection; provide a quantitative measure (e.g., variance or quartiles of XCOMET scores within coverage bins) to support the conclusion.
  4. [§5.2] 'All evaluations are conducted in a zero-shot setting' is ambiguous; M2PO is fine-tuned on a preference dataset. Clarify that zero-shot refers to test-domain generalization rather than absence of task fine-tuning.
  5. [Table 4] Clarify what 'Orig.' means for each DPO-like algorithm (same preference data? same hyperparameters?) and whether the +M2PO setting uses the same multi-pair data for all algorithms.
  6. [Appendix A] Provide the full coverage-score prompt and the exact computation of S_align (e.g., is it an alignment coverage ratio?); the current one-sentence description is insufficient for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: M2PO's quality claims rest on external COMET/XCOMET evaluation; the faithfulness-metric overlap is a validity concern, not an equation-level reduction.

full rationale

The derivation chain does not reduce to its inputs. The training reward is rs = r_qe + lambda_f * S_align (Eq. 2), with r_qe from KIWI-XXL and S_align from WSPAlign; the main translation-quality evaluations use COMET-22 and XCOMET, and the paper explicitly states a "strict no-leakage principle where the training reward metric (KIWI-XXL) is never used for evaluation." Thus the headline quality claim is not a fitted-input prediction. The faithfulness evaluation does use a Gemini-2.0-Flash Coverage Score that is semantically similar to the WSPAlign-based factuality bonus, but the paper does not equate them: the reward bonus is a word-alignment score, while the evaluation is an LLM-judge coverage percentage with a different prompt. No equation or construction step makes the evaluated metric identical to the optimized reward, so there is no reduction by construction. The motivational analysis in Section 3 is anchored to the external HalOmi human-labeled benchmark, not to the method's own scores. The useful references to WSPAlign and coverage-style scoring are to Wu et al. (2023, 2024), not to the present authors, so there is no load-bearing self-citation chain or imported uniqueness theorem. The main legitimate concern is external validity: WSPAlign and the Gemini coverage judge are not validated against human faithfulness judgments, and no human evaluation is reported. That is a correctness/evidence issue, not circularity. Accordingly the circularity score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 7 axioms · 0 invented entities

The paper is benchmark-driven, so the free parameters are conventional hyperparameters, but none are justified by sensitivity analysis or held-out selection. The central risk is the unvalidated faithfulness signals (S_align and Gemini coverage) that are load-bearing for both the reward and the evaluation; this is where the circularity and reproducibility concerns concentrate.

free parameters (7)
  • λ_f = 1.0
    Weight for WSPAlign factuality bonus in Eq. (2); chosen by hand, no sensitivity analysis reported.
  • β = 0.1
    DPO temperature in Eq. (5); hand-set.
  • α_t schedule = linear 0.1 → 0.9
    Curriculum weight in Eq. (3) fusing static and dynamic scores; chosen by hand, no ablation of the schedule.
  • τ_s = 0.4
    Temperature for ListNet distributions in Eqs. (6-7).
  • τ_w = 0.3
    Temperature for pair weights w_i in the multi-pair DPO loss.
  • λ_pref, λ_rank, λ_bc = 1.0, 0.5, 1.0
    Loss weights in Eq. (4); no tuning or sensitivity study shown.
  • K = 16
    Candidate pool size before pair construction; fixed but arbitrary.
axioms (7)
  • standard math Bradley-Terry model and the DPO reward/policy equivalence hold for the pair construction
    Invoked in §4.1 via DPO (Rafailov et al.), but Eq. (5) then omits the reference-model term.
  • domain assumption KIWI-XXL QE scores correlate with human translation quality
    Used as the base quality score in §4.3 Eq. (2); the paper's own motivation says QE is unreliable for partial errors.
  • domain assumption WSPAlign word alignment can be converted into a 0-100 factuality bonus that penalizes hallucinations
    This conversion is not specified in §4.3 or Appendix A; S_align is central to the static reward.
  • ad hoc to paper Gemini-2.0-Flash 'coverage score' is a reliable gold standard for faithfulness and hallucination
    Used both to motivate the QE blind spot (§3, Table 1) and as the final faithfulness evaluation (§5.2); no human calibration is provided.
  • domain assumption The model's own log-probability can act as an online reward without reward hacking
    r_d = log π_θ enters the fused dynamic score in Eq. (3); the paper acknowledges reward hacking risk but does not measure it.
  • domain assumption The candidate pool from the SFT model and GPT-4o-mini, augmented with mined hard negatives, spans the useful quality spectrum
    Stage 1 (§4.2) relies on this diversity; no analysis of candidate distribution is given.
  • domain assumption The correlation increase in Table 1 after deleting partial-error examples indicates a QE blind spot rather than a distributional artifact
    The conclusion in §3 is based on this re-computation without significance testing.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of $M^2PO$: Multi-Perspective Multi-Pair Preference Optimization for Machine Translation." pith.science (2026). https://pith.science/paper/55ILPIW2

@misc{pith2026251013434,
  author       = {Pith},
  title        = {Pith review of: $M^2PO$: Multi-Perspective Multi-Pair Preference Optimization for Machine Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/55ILPIW2}},
  note         = {Machine review of arXiv:2510.13434}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Aligning Large Language Models (LLMs) with human preferences is pivotal for Machine Translation (MT), yet current approaches are often hindered by misleading reward signals. Our analysis reveals that prevailing Quality Estimation (QE) models exhibit a systematic blind spot toward partial errors, specifically partial hallucinations and omissions, often favoring superficially fluent but unfaithful translations. To address this issue, we propose $M^2PO$ (Multi-Perspective Multi-Pair Preference Optimization), a data-centric framework for preference optimization in machine translation. First, to correct the bias toward fluency, $M^2PO$ uses a dual-perspective mechanism that decouples semantic fidelity from fluency and prioritizes faithfulness through a curriculum strategy. Second, after correcting this bias, partial errors fall between perfect and severely incorrect translations, making them difficult to learn through standard best-versus-worst comparisons. We therefore introduce a multi-pair objective that leverages the full candidate list to capture these fine-grained error signals. Experiments on WMT23, WMT24, and FLORES-200 show that $M^2PO$ enables a 9B model to outperform leading open-source baselines and achieve parity with proprietary models such as GPT-4o and Gemini-2.0-Flash, demonstrating strong potential for efficient and high-fidelity LLM-based translation. Our code and dataset will be released.

Figures

Figures reproduced from arXiv: 2510.13434 by Bo Zeng, Hao Wang, Heng Liu, Jiang Zhou, Liangying Shao, Linlong Xu, Longyue Wang, Tianyu Dong, Weihua Luo, Xiangxiang Zeng, Xiaohu Zhao, Xinwei Wu, Yangyang Liu, Yichen Dong.

Figure 1
Figure 1. Figure 1: The relationship between the XCOMET score [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The M2PO framework. Stage 1: Candidate Set Construction creates a diverse dataset. Stage 2: Dynamic Preference Modeling computes a robust static score (rs) offline by combining QE with a hallucination penalty, then fuses it with the model’s own judgment (rd) to create an online, dynamic fused score (rfused) for ranking. Stage 3: Multi-Component Optimization uses this ranking to form multiple preference pai… view at source ↗
Figure 3
Figure 3. Figure 3: Translation Quality (XCOMET) vs. Faithful [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: displays the prompts used to ensure repli￾cable translation and evaluation. Our prompt for Coverage Calculation is adapted from (Wu et al., 2024). For proprietary models, we slightly modify the prompt from (Xu et al., 2024b) to align with the ALMA format. Our M2PO model and all ALMA baselines use the standard prompt from the original ALMA study (Xu et al., 2023). Given a source sentence {source_text} in {s… 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

44 extracted references · 31 linked inside Pith

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    Pith/arXiv arXiv 2023

  2. [2]

    Sweta Agrawal, Jos \'e GC De Souza, Ricardo Rei, Ant \'o nio Farinhas, Gon c alo Faria, Patrick Fernandes, Nuno M Guerreiro, and Andre Martins. 2024. Modeling user preferences with automatic metrics: creating a high-quality preference dataset for machine translation. arXiv preprint arXiv:2410.07779

    Pith/arXiv arXiv 2024

  3. [3]

    Duarte M Alves, Jos \'e Pombal, Nuno M Guerreiro, Pedro H Martins, Jo \ a o Alves, Amin Farajian, Ben Peters, Ricardo Rei, Patrick Fernandes, Sweta Agrawal, and 1 others. 2024. Tower: An open multilingual large language model for translation-related tasks. arXiv preprint arXiv:2402.17733

    Pith/arXiv arXiv 2024

  4. [4]

    Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007. Learning to rank: from pairwise approach to listwise approach. In Proceedings of the 24th international conference on Machine learning, pages 129--136

    2007

  5. [5]

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261

    Pith/arXiv arXiv 2025

  6. [6]

    Marta R Costa-Juss \`a , James Cross, Onur C elebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, and 1 others. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672

    Pith/arXiv arXiv 2022

  7. [7]

    Guofeng Cui, Pichao Wang, Yang Liu, Zemian Ke, Zhu Liu, and Vimal Bhat. 2025. Crpo: Confidence-reward driven preference optimization for machine translation. arXiv preprint arXiv:2501.13927

    Pith/arXiv arXiv 2025

  8. [8]

    David Dale, Elena Voita, Lo \" c Barrault, and Marta R Costa-Juss \`a . 2022. Detecting and mitigating hallucinations in machine translation: Model internal workings alone do well, sentence similarity even better. arXiv preprint arXiv:2212.08597

    Pith/arXiv arXiv 2022

  9. [9]

    David Dale, Elena Voita, Janice Lam, Prangthip Hansanti, Christophe Ropers, Elahe Kalbassi, Cynthia Gao, Lo \" c Barrault, and Marta R Costa-juss \`a . 2023. Halomi: A manually annotated benchmark for multilingual hallucination and omission detection in machine translation. arXiv preprint arXiv:2305.11746

    Pith/arXiv arXiv 2023

  10. [10]

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306

    Pith/arXiv arXiv 2024

  11. [11]

    Evangelia Gogoulou, Shorouq Zahra, Liane Guillou, Luise D \"u rlich, and Joakim Nivre. 2025. Can llms detect intrinsic hallucinations in paraphrasing and machine translation? arXiv preprint arXiv:2504.20699

    Pith/arXiv arXiv 2025

  12. [12]

    Nuno M Guerreiro, Duarte M Alves, Jonas Waldendorf, Barry Haddow, Alexandra Birch, Pierre Colombo, and Andr \'e FT Martins. 2023. Hallucinations in large multilingual translation models. Transactions of the Association for Computational Linguistics, 11:1500--1517

    2023

  13. [13]

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

    2024

  14. [14]

    Nuno M Guerreiro, Elena Voita, and Andr \'e FT Martins. 2022. Looking for a needle in a haystack: A comprehensive study of hallucinations in neural machine translation. arXiv preprint arXiv:2208.05309

    Pith/arXiv arXiv 2022

  15. [15]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    Pith/arXiv arXiv 2025

  16. [16]

    Zhiwei He, Xing Wang, Wenxiang Jiao, Zhuosheng Zhang, Rui Wang, Shuming Shi, and Zhaopeng Tu. 2024. Improving machine translation with human feedback: An exploration of quality estimation as a reward model. arXiv preprint arXiv:2401.12873

    Pith/arXiv arXiv 2024

  17. [17]

    Joey Hejna, Rafael Rafailov, Harshit Sikchi, Chelsea Finn, Scott Niekum, W Bradley Knox, and Dorsa Sadigh. 2023. Contrastive preference learning: learning from human feedback without rl. arXiv preprint arXiv:2310.13639

    Pith/arXiv arXiv 2023

  18. [18]

    Amr Hendy, Mohamed Abdelrehim, Amr Sharaf, Vikas Raunak, Mohamed Gabr, Hitokazu Matsushita, Young Jin Kim, Mohamed Afify, and Hany Hassan Awadalla. 2023. How good are gpt models at machine translation? a comprehensive evaluation. arXiv preprint arXiv:2302.09210

    Pith/arXiv arXiv 2023

  19. [19]

    Jiwoo Hong, Noah Lee, and James Thorne. 2024. Orpo: Monolithic preference optimization without reference model. arXiv preprint arXiv:2403.07691

    Pith/arXiv arXiv 2024

  20. [20]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    Pith/arXiv arXiv 2024

  21. [21]

    Wenxiang Jiao, Wenxuan Wang, Jen-tse Huang, Xing Wang, and Zhaopeng Tu. 2023. Is chatgpt a good translator? a preliminary study. arXiv preprint arXiv:2301.08745, 1(10)

    Pith/arXiv arXiv 2023

  22. [22]

    Tom Kocmi, Eleftherios Avramidis, Rachel Bawden, Ond r ej Bojar, Anton Dvorkovich, Christian Federmann, Mark Fishel, Markus Freitag, Thamme Gowda, Roman Grundkiewicz, and 1 others. 2024. Findings of the wmt24 general machine translation shared task: The llm era is here but mt is not solved yet. In Proceedings of the Ninth Conference on Machine Translation...

    2024

  23. [23]

    Yeskendir Koishekenov, Alexandre Berard, and Vassilina Nikoulina. 2022. Memory-efficient nllb-200: Language-specific expert pruning of a massively multilingual machine translation model. arXiv preprint arXiv:2212.09811

    Pith/arXiv arXiv 2022

  24. [24]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

    Pith/arXiv arXiv 2024

  25. [25]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101

    Pith/arXiv arXiv 2017

  26. [26]

    Yinquan Lu, Wenhao Zhu, Lei Li, Yu Qiao, and Fei Yuan. 2024. Llamax: Scaling linguistic horizons of llm by enhancing translation capabilities beyond 100 languages. arXiv preprint arXiv:2407.05975

    Pith/arXiv arXiv 2024

  27. [27]

    Yu Meng, Mengzhou Xia, and Danqi Chen. 2024. Simpo: Simple preference optimization with a reference-free reward. Advances in Neural Information Processing Systems, 37:124198--124235

    2024

  28. [28]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730--27744

    2022

  29. [29]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728--53741

    2023

  30. [30]

    Ricardo Rei, Nuno M Guerreiro, Jos \'e Pombal, Jo \ a o Alves, Pedro Teixeirinha, Amin Farajian, and Andr \'e FT Martins. 2025. Tower+: Bridging generality and translation specialization in multilingual llms. arXiv preprint arXiv:2506.17080

    Pith/arXiv arXiv 2025

  31. [31]

    Haoxiang Sun, Ruize Gao, Pei Zhang, Baosong Yang, and Rui Wang. 2025. Enhancing machine translation with self-supervised preference data. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 23916--23934

    2025

  32. [32]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    Pith/arXiv arXiv 2023

  33. [33]

    Jannis Vamvas and Rico Sennrich. 2022. As little as possible, as much as necessary: Detecting over-and undertranslations with contrastive conditioning. arXiv preprint arXiv:2203.01927

    Pith/arXiv arXiv 2022

  34. [34]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  35. [35]

    Qiyu Wu, Masaaki Nagata, Zhongtao Miao, and Yoshimasa Tsuruoka. 2024. Word alignment as preference for machine translation. arXiv preprint arXiv:2405.09223

    Pith/arXiv arXiv 2024

  36. [36]

    Qiyu Wu, Masaaki Nagata, and Yoshimasa Tsuruoka. 2023. Wspalign: Word alignment pre-training via large-scale weakly supervised span prediction. arXiv preprint arXiv:2306.05644

    Pith/arXiv arXiv 2023

  37. [37]

    Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. 2023. A paradigm shift in machine translation: Boosting translation performance of large language models. arXiv preprint arXiv:2309.11674

    Pith/arXiv arXiv 2023

  38. [38]

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

    Pith/arXiv arXiv 2024

  39. [39]

    Haoran Xu, Amr Sharaf, Yunmo Chen, Weiting Tan, Lingfeng Shen, Benjamin Van Durme, Kenton Murray, and Young Jin Kim. 2024 b . Contrastive preference optimization: Pushing the boundaries of llm performance in machine translation. arXiv preprint arXiv:2401.08417

    Pith/arXiv arXiv 2024

  40. [40]

    Zonghan Yang, Yong Cheng, Yang Liu, and Maosong Sun. 2019. Reducing word omission errors in neural machine translation: A contrastive learning approach. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6191--6196

    2019

  41. [41]

    Jiali Zeng, Fandong Meng, Yongjing Yin, and Jie Zhou. 2024. Teaching large language models to translate with comparison. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19488--19496

    2024

  42. [42]

    Shaolei Zhang, Qingkai Fang, Zhuocheng Zhang, Zhengrui Ma, Yan Zhou, Langlin Huang, Mengyu Bu, Shangtong Gui, Yunji Chen, Xilin Chen, and 1 others. 2023. Bayling: Bridging cross-lingual alignment and instruction following through interactive translation for large language models. arXiv preprint arXiv:2306.10968

    Pith/arXiv arXiv 2023

  43. [43]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  44. [44]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.