Pith. sign in

REVIEW 3 major objections 5 minor 24 references

MBR decoding decomposed into four probabilistic channels

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 · glm-5.2

2026-07-07 23:50 UTC pith:ARY7SERY

load-bearing objection Noisy-channel decomposition of MBR decoding is a clean interpretive framework; the improvement claim is thin the 3 major comments →

arxiv 2607.05198 v1 pith:ARY7SERY submitted 2026-07-06 cs.LG cs.AIcs.CL

Noisy-Channel Minimum Bayes Risk Decoding

Yusuke Sakai , Hidetaka Kamigaito , Taro Watanabe This is my paper
classification cs.LG cs.AIcs.CL
keywords Minimum Bayes Risk decodingnoisy channel modeltext generationevaluation metricsmachine translationsummarizationimage captioningBayes rule decomposition
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.

Minimum Bayes Risk (MBR) decoding is a method for selecting the best text output from a set of candidates by picking the one that maximizes expected similarity to sampled pseudo-references, rather than simply picking the most probable output. This paper shows that the scoring formula underlying MBR decoding can be algebraically decomposed into four probabilistic components — the hypothesis-to-reference likelihood, the reference-to-hypothesis likelihood, the hypothesis prior, and the reference prior — using a noisy-channel (Bayes-rule) factorization. By introducing adjustable weights on each component, the authors create a unified framework that subsumes several existing MBR variants as special cases of four hyperparameters, and they show empirically that the contribution of each channel depends strongly on which evaluation metric (BLEU, chrF, COMET, or BERTScore) is used as the utility function, but is largely consistent across tasks (translation, summarization, captioning). The paper also demonstrates that tuning the channel weights can yield measurable improvements over standard MBR decoding.

Core claim

The central discovery is that the MBR scoring function, which selects a hypothesis by averaging an evaluation metric score over pseudo-references, can be rewritten through Bayes-rule factorization as a product of four terms: P(r_j|h_i) (how likely the reference is given the hypothesis), P(h_i) (the hypothesis prior), P(h_i|r_j) (how likely the hypothesis is given the reference), and P(r_j) (the reference prior). The inverse-direction likelihood P(r_j|h_i) is computed by simply swapping the arguments of the metric function — that is, evaluating f_θ(r_j, h_i) instead of f_θ(h_i, r_j). This decomposition reveals that different evaluation metrics implicitly weight these four channels in distinct

What carries the argument

Noisy-channel decomposition of MBR scoring into four weighted probabilistic terms (Eq. 13): argmax over h_i of the sum over r_j of sqrt(P(r_j|h_i)^alpha * P(h_i)^beta * P(h_i|r_j)^gamma * P(r_j)^delta). Existing MBR variants (MAP, original MBR, conditional/BOC, swap, inverse) correspond to specific settings of (alpha, beta, gamma, delta).

Load-bearing premise

The paper assumes that an evaluation metric score computed in the reverse direction — f_θ(r_j, h_i) instead of f_θ(h_i, r_j) — can serve as a valid stand-in for the probability P(r_j|h_i). This treats an asymmetric scoring function as if it were a likelihood, which is the step that enables the entire bidirectional analysis.

What would settle it

If the inverse-direction metric score f_θ(r_j, h_i) bears no meaningful relationship to an actual probability distribution over references given hypotheses, then the four-term decomposition is a mathematical rearrangement without probabilistic grounding, and the channel weights would be tuning arbitrary score combinations rather than meaningful probabilistic components.

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

If this is right

  • Practitioners can tune four channel weights to improve MBR decoding performance for a given evaluation metric, with the paper showing that optimal weights transfer across tasks when the metric is held fixed.
  • The decomposition explains why model-based MBR (which explicitly incorporates model probabilities as the reference prior) works well for BLEU and BERTScore but less stably for chrF and COMET — the reference prior channel has metric-dependent sensitivity.
  • The hypothesis-to-reference likelihood P(r_j|h_i), computed by inverting the metric direction, is identified as a previously underexplored factor that contributes meaningfully to MBR performance, especially for symmetric metrics like BERTScore.
  • The framework provides a diagnostic tool: by varying one weight at a time, one can characterize how a given evaluation metric interacts with each probabilistic component, enabling more principled metric selection for MBR decoding.

Where Pith is reading between the lines

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

  • If the channel-weight trends are truly metric-specific and task-agnostic, then one could precompute optimal weight profiles per metric and deploy them as default settings, removing the need for per-task hyperparameter search.
  • The finding that BERTScore benefits from high weighting on the hypothesis-to-reference channel (alpha) while BLEU and chrF prefer low alpha suggests that symmetric metrics and asymmetric metrics engage the bidirectional decomposition in fundamentally different ways — a property that could guide the design of new utility functions specifically tailored for MBR decoding.
  • The framework could be extended to learn the four channel weights automatically (e.g., via Bayesian optimization or differentiable tuning), rather than relying on grid search, which would make the method more practical for production systems.
  • The decomposition may have implications for understanding why MBR decoding sometimes underperforms MAP decoding: if the implicit weighting of the four channels is suboptimal for a given metric-task combination, the noisy-channel formulation could diagnose and correct the mismatch.

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 / 5 minor

Summary. This paper introduces a noisy-channel decomposition of Minimum Bayes Risk (MBR) decoding, reformulating the MBR scoring function into four weighted probabilistic components: the hypothesis-to-reference likelihood P(r_j|h_i), the hypothesis prior P(h_i), the reference-to-hypothesis likelihood P(h_i|r_j), and the reference prior P(r_j). The decomposition (Eq. 13) generalizes standard MBR (which corresponds to weights (α,β,γ,δ)=(0,0,1,1)) and encompasses MAP decoding, conditional variants, and model-based MBR as special cases (Table 1). The authors conduct a grid search over the four-dimensional weight space across 12 machine translation scenarios (WMT22/23, 3 language pairs, both directions) plus summarization (CNN/DM, XSum, SAMSum) and captioning (MSCOCO, NoCaps) tasks, using BLEU, chrF, COMET, and BERTScore as utility functions. The analysis examines whether channel-wise importance is metric-specific (RQ1) and task-agnostic (RQ2), finding that trends are primarily driven by the choice of utility function rather than the task. The paper concludes that appropriate channel weighting can yield improvements over standard MBR, supported by a transfer experiment from WMT22 En-De to WMT23 En-De.

Significance. The paper provides a useful interpretive framework for understanding MBR decoding variants through a unified noisy-channel lens. The unification in Table 1, showing that MAP, original MBR, conditional/BOC variants, and swap configurations are all special cases of the four-parameter family, is a clean conceptual contribution. The experimental scope is broad, spanning 12 MT scenarios plus summarization and captioning with multiple utility functions. The observation that channel importance is metric-specific but largely task-agnostic (for BERTScore) is a practically useful finding. The framework also provides a theoretical explanation for why model-based MBR (Jinnai et al., 2024) benefits BLEU and BERTScore but not chrF or COMET, which is a nice connection between theory and prior empirical results.

major comments (3)
  1. §5 (Conclusion), transfer experiment: The claim that 'reweighting can yield performance improvements over standard MBR' is supported by a single transfer result: best weights from WMT22 En-De applied to WMT23 En-De, yielding 50.05 to 50.18 chrF (0.13 absolute). This is the only piece of evidence for the improvement claim, and it is presented without significance testing, confidence intervals, or multiple seeds. Given that the grid search explores 9^4 = 6561 configurations on the same evaluation metric used as the utility function, the overfitting risk is substantial. A 0.13 chrF difference is well within the sampling noise expected for MBR with 256 candidates. This is load-bearing for the paper's claim of 'improvements over original MBR decoding' and needs either (a) significance testing (e.g., paired bootstrap or approximate randomization), (b) multiple transfer directions (not just En-
  2. §3.1, Eqs. (8)-(11): The derivation chain from Eq. (8) to Eq. (11) contains an algebraic inconsistency. Eq. (8) is argmax sum_j P(h_i|r_j) P(r_j). Eq. (9) writes this as argmax sum_j sqrt(P(h_i|r_j)^2 * P(r_j)), which equals argmax sum_j P(h_i|r_j) * sqrt(P(r_j)), NOT P(h_i|r_j)*P(r_j) as in Eq. (8). For the equality to hold, Eq. (9) should contain P(r_j)^2 inside the square root: sqrt(P(h_i|r_j)^2 * P(r_j)^2) = P(h_i|r_j)*P(r_j). The final result in Eq. (11) is correct in form, but the intermediate step in Eq. (9) has a typo that breaks the chain. This should be corrected.
  3. §3.1, Eq. (12) and surrounding text: The assumption that f_theta(r_j, h_i) / sum_j' f_theta(r_j', h_i) serves as a proxy for P(r_j|h_i) is the load-bearing premise enabling the bidirectional analysis. The paper acknowledges this is 'decision-theoretic rather than a true probabilistic decomposition,' which is an appropriate caveat. However, for asymmetric metrics (BLEU, COMET), P(r_j|h_i) estimated from f_theta(r_j,h_i) and P(h_i|r_j) estimated from f_theta(h_i,r_j) derive from different scoring functions applied in different directions, so they do not correspond to a single joint distribution P(h_i, r_j). Bayes' rule (invoked in the step from Eq. 8 to Eq. 9-10) does not formally connect these two quantities when they come from asymmetric metrics. The paper should more explicitly state the scope of the probabilistic interpretation: the decomposition is exact when metrics are symmetric (e
minor comments (5)
  1. Figures 1-3: The z-score normalization makes it difficult to assess absolute performance differences. Including a table with raw scores for key configurations (at least standard MBR and best-found weights) would help readers gauge practical significance.
  2. §4.1.1: The grid search uses 9 values per parameter across 4 parameters (6561 configs) on 12 MT scenarios. It would help to state the total computational cost and whether the grid was coarser for the summarization/captioning experiments.
  3. §3.2, Table 1: The 'Ours (NEW)' column lists (α,β,γ,δ)=(1,0,0,0), but it is unclear what specific configuration this refers to or how it was selected. Clarification would help.
  4. The paper mentions code will be released 'upon completion' of a repository consolidation. For reproducibility, providing the grid search configurations and best weights as a supplementary table would be valuable.
  5. §4.1.2, Finding 2: The claim that BERTScore benefits from high α because it 'computes an F-score based on symmetric token-level alignment' could be stated more precisely — BERTScore's symmetry means the swap direction is identical to the original, so high α effectively amplifies the same signal as γ.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive report. The referee identifies three major issues: (1) the transfer experiment supporting our 'improvements over standard MBR' claim lacks significance testing and multiple transfer directions; (2) an algebraic typo in the derivation chain from Eq. (8) to Eq. (11); and (3) the scope of the probabilistic interpretation needs to be stated more explicitly, particularly for asymmetric metrics where the two directional quantities do not correspond to a single joint distribution. We agree with all three points and will address each in the revision.

read point-by-point responses

  1. Referee: §5, transfer experiment: The claim that 'reweighting can yield performance improvements over standard MBR' is supported by a single transfer result (WMT22 En-De to WMT23 En-De, 50.05 to 50.18 chrF, 0.13 absolute) without significance testing, confidence intervals, or multiple seeds. Overfitting risk is substantial given 9^4=6561 configurations. A 0.13 chrF difference is within sampling noise for MBR with 256 candidates.

    Authors: The referee is correct on all counts. The single transfer result without significance testing is insufficient to support the improvement claim, and the overfitting concern is well-founded given the size of the search space relative to the single transfer evaluation. We will address this through the following revisions: (a) we will add paired bootstrap resampling significance tests for the transfer experiment and report confidence intervals; (b) we will add at least two additional transfer directions (e.g., WMT22 De-En to WMT23 De-En, and a transfer from MT to summarization or captioning under BERTScore) to demonstrate that the finding is not specific to one language pair; (c) we will add explicit discussion of the overfitting risk inherent in the grid search and frame the transfer result as suggestive rather than definitive. If the additional experiments do not yield statistically significant improvements, we will accordingly soften the claim in the abstract and conclusion from 'may lead to improvements' to a more conservative statement about the potential for improvements subject to further validation. The primary contribution of the paper is the interpretive framework and the analysis of channel-wise importance; the improvement claim is secondary and we will present it with appropriate caution. revision: yes

  2. Referee: §3.1, Eqs. (8)-(11): The derivation chain contains an algebraic inconsistency. Eq. (9) writes sqrt(P(h_i|r_j)^2 * P(r_j)), which equals P(h_i|r_j)*sqrt(P(r_j)), NOT P(h_i|r_j)*P(r_j) as in Eq. (8). For the equality to hold, Eq. (9) should contain P(r_j)^2 inside the square root.

    Authors: The referee is correct. Equation (9) should read sqrt(P(h_i|r_j)^2 * P(r_j)^2) to maintain equality with Equation (8). As the referee notes, the final result in Equation (11) is correct in form; only the intermediate step in Equation (9) contains the typo. We will correct this in the revised manuscript. revision: yes

  3. Referee: §3.1, Eq. (12) and surrounding text: For asymmetric metrics, P(r_j|h_i) estimated from f_theta(r_j,h_i) and P(h_i|r_j) estimated from f_theta(h_i,r_j) derive from different scoring functions applied in different directions, so they do not correspond to a single joint distribution P(h_i, r_j). Bayes' rule does not formally connect these two quantities when they come from asymmetric metrics. The paper should more explicitly state the scope of the probabilistic interpretation.

    Authors: The referee raises a valid and important point about the scope of the probabilistic interpretation. We agree that for asymmetric metrics, the two directional quantities P(r_j|h_i) and P(h_i|r_j) as estimated from f_theta(r_j, h_i) and f_theta(h_i, r_j) respectively do not arise from a single joint distribution, and therefore the application of Bayes' rule in the step from Eq. (8) to Eqs. (9)-(10) is not formally justified in the asymmetric case. We will revise the manuscript to state this explicitly: the decomposition is exact when the utility function is symmetric (e.g., chrF, BERTScore), where f_theta(r_j, h_i) = f_theta(h_i, r_j) and the two directional quantities do correspond to a single joint distribution. For asymmetric metrics (e.g., BLEU, COMET), the decomposition should be understood as a decision-theoretic heuristic inspired by the noisy-channel framework rather than a true probabilistic decomposition. We will add this clarification to Section 3.1, immediately following the introduction of Eq. (12), and adjust the framing in the abstract and introduction to avoid overstating the probabilistic interpretation for asymmetric metrics. The existing caveat ('decision-theoretic concept rather than a true probabilistic decomposition') will be expanded with this specific scope limitation. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained, with one minor self-citation for a supporting interpretation.

full rationale

The paper's core derivation (Eqs. 4→13) is a genuine algebraic reparametrization of MBR scoring, not a circular restatement. Each of the four decomposed components is a distinct quantity: P(h_i|r_j) is the normalized forward metric score, P(r_j|h_i) is the normalized swapped metric score (Eq. 12), and P(h_i), P(r_j) are model priors. Table 1 shows the decomposition reduces to standard MBR at (α,β,γ,δ)=(0,0,1,1), confirming it is a generalization rather than a tautology. The 'improvement' claim (50.05→50.18 chrF, WMT22→WMT23 transfer) involves grid-searching weights on one dataset and evaluating on another—this is standard hyperparameter transfer with overfitting risk, but not circular by construction since the test set is different from the tuning set. The one self-citation is Kamigaito et al. (2025), co-authored by a present author, invoked for the Bayes Optimal Classifier interpretation (Eq. 16-17); this is a minor supporting point, not load-bearing for the main decomposition. The algebraic issue in Eq. 9 (P(r_j) should be P(r_j)² for the equality chain to hold) is a correctness concern, not a circularity concern—the final result (Eq. 11) is what the paper actually uses and is algebraically correct. No step in the derivation chain reduces to its own inputs by definition or by self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 2 axioms · 0 invented entities

The framework introduces four free hyperparameters (α, β, γ, δ) that are grid-searched to find optimal weighting. The key ad-hoc axiom is that swapping metric arguments yields a valid likelihood proxy (Eq. 12). No new physical or abstract entities are invented.

free parameters (4)
  • α = grid-searched over [0.00, 0.25, ..., 2.00]
    Weight for P(r_j|h_i), tuned to data
  • β = grid-searched over [0.00, 0.25, ..., 2.00]
    Weight for P(h_i), tuned to data
  • γ = grid-searched over [0.00, 0.25, ..., 2.00]
    Weight for P(h_i|r_j), tuned to data
  • δ = grid-searched over [0.00, 0.25, ..., 2.00]
    Weight for P(r_j), tuned to data
axioms (2)
  • ad hoc to paper f_θ(r_j, h_i) / Σ f_θ(r_j', h_i) is a valid proxy for P(r_j|h_i)
    Eq. 12 assumes an asymmetric evaluation metric can be normalized to serve as a likelihood, which is the core enabling premise for the bidirectional analysis.
  • domain assumption Utility functions can be treated as non-negative
    Footnote 3 states this is valid because argmax is invariant to additive shifts, used in deriving Eq. 6.

pith-pipeline@v1.1.0-glm · 25246 in / 1734 out tokens · 101425 ms · 2026-07-07T23:50:03.146419+00:00 · methodology

0 comments
read the original abstract

Minimum Bayes Risk (MBR) decoding yields more robust and higher-quality text generation than maximum a posteriori (MAP) decoding by selecting hypotheses that maximize expected utility over sampled pseudo-references. However, there exists a discrepancy in the design: hypothesis selection calculates expected utility scores conditioned on given pseudo-references, while commonly used evaluation metrics, e.g., BLEU and COMET, are asymmetric. Therefore, it is important to consider both hypothesis-to-reference and reference-to-hypothesis directional effects. In this study, we introduce a noisy channel decomposition of MBR decoding that naturally incorporates bidirectional effects to account for these asymmetries. We decompose MBR decoding into four interacting components: hypothesis-to-reference likelihood, reference-to-hypothesis likelihood, hypothesis prior, and reference prior. This decomposition provides a unified interpretation of existing MBR variants and enables metric- and task-specific interpretability by isolating the contribution of each channel. Our comprehensive analysis reveals that channel-wise contributions exhibit distinct characteristics across metrics while remaining consistent across tasks, and suggests that appropriate channel weighting may lead to improvements over original MBR decoding.

Figures

Figures reproduced from arXiv: 2607.05198 by Hidetaka Kamigaito, Taro Watanabe, Yusuke Sakai.

Figure 1
Figure 1. Figure 1: Results on machine translation tasks across 12 evaluation scenarios spanning language directions and datasets. Rows correspond to evaluation metrics, and columns correspond to the term weights α, β, γ, and δ. Each plot reports the result obtained by grid searching over [0.00, 0.25, 0.50, 0.75, 1.00, 1.25, 1.50, 1.75, 2.00] for the corresponding term weight. We report z-scores normalized with respect to the… view at source ↗
Figure 2
Figure 2. Figure 2: Results on summarization, captioning, and machine translation tasks with BERTScore. The machine translation results are taken from [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Results on captioning and machine translation tasks with BLEU. The machine translation results are taken from [PITH_FULL_IMAGE:figures/full_fig_p009_3.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

24 extracted references · 24 canonical work pages · 1 internal anchor

  1. [1]

    bigpicture-1.9/

    URL https://aclanthology.org/2023. bigpicture-1.9/. Cheng, J. and Vlachos, A. Faster minimum Bayes risk decoding with confidence-based pruning. In Bouamor, H., Pino, J., and Bali, K. (eds.),Proceed- ings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 12473–12480, Singapore, December 2023. Association for Computa- tional Li...

    work page doi:10.18653/v1/2023.emnlp-main 2023

  2. [3]

    emnlp-main.1710/

    URL https://aclanthology.org/2025. emnlp-main.1710/. Deguchi, H., Imamura, K., Nishida, Y ., Sakai, Y ., Vasselli, J., and Watanabe, T. NAIST-NICT WMT’23 general MT task submission. In Koehn, P., Haddow, B., Kocmi, T., and Monz, C. (eds.),Proceedings of the Eighth Confer- ence on Machine Translation, pp. 110–118, Singapore, December 2023. Association for ...

    work page doi:10.18653/v1/2023.wmt-1.7 2025

  3. [4]

    coling-main.398/

    URL https://aclanthology.org/2020. coling-main.398/. Eikema, B. and Aziz, W. Sampling-based approxima- tions to minimum Bayes risk decoding for neural ma- chine translation. In Goldberg, Y ., Kozareva, Z., and Zhang, Y . (eds.),Proceedings of the 2022 Confer- ence on Empirical Methods in Natural Language Pro- cessing, pp. 10978–10993, Abu Dhabi, United Ar...

    work page doi:10.18653/v1/2022.emnlp-main 2020

  4. [5]

    naacl-long.331/

    URL https://aclanthology.org/2022. emnlp-main.754/. Eikema, B., Rutkiewicz, A., and Giulianelli, M. Structure- conditional minimum Bayes risk decoding. In Christodoulopoulos, C., Chakraborty, T., Rose, C., and Peng, V . (eds.),Proceedings of the 2025 Conference on Empirical Methods in Natural Language Process- ing, pp. 31706–31723, Suzhou, China, November...

    work page doi:10.18653/v1/2025.emnlp-main 2022

  5. [6]

    Beyond English-Centric Multilingual Machine Translation

    URL https://aclanthology.org/2025. emnlp-main.1616/. Fan, A., Bhosale, S., Schwenk, H., Ma, Z., El-Kishky, A., Goyal, S., Baines, M., Celebi, O., Wenzek, G., Chaudhary, V ., Goyal, N., Birch, T., Liptchinsky, V ., Edunov, S., Grave, E., Auli, M., and Joulin, A. Beyond english- centric multilingual machine translation, 2020. URL https://arxiv.org/abs/2010....

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.18653/v1/2024.wmt-1.126 2025

  6. [7]

    findings-emnlp.617/

    URL https://aclanthology.org/2023. findings-emnlp.617/. Freitag, M., Mathur, N., Deutsch, D., Lo, C.-K., Avramidis, E., Rei, R., Thompson, B., Blain, F., Kocmi, T., Wang, J., Adelani, D. I., Buchicchio, M., Zerva, C., and Lavie, A. Are LLMs breaking MT metrics? results of the WMT24 metrics shared task. In Haddow, B., Kocmi, T., Koehn, P., and Monz, C. (ed...

    work page doi:10.18653/v1/2024.wmt-1.2 2023

  7. [8]

    URL https://www.sciencedirect.com/ science/article/pii/S0885230800901384

    doi: https://doi.org/10.1006/csla.2000.0138. URL https://www.sciencedirect.com/ science/article/pii/S0885230800901384. Hayakawa, A., Khallaf, N., Saggion, H., and Sharoff, S. UoL-UPF at TSAR 2025 shared task a generate-and- select approach for readability-controlled text simplifi- cation. In Shardlow, M., Alva-Manchego, F., North, K., Stodden, R., Saggion...

    work page doi:10.1006/csla.2000.0138 2000

  8. [9]

    Ichihara, Y ., Jinnai, Y ., Ariu, K., Morimura, T., and Uchibe, E

    URL https://openreview.net/forum? id=XPZIaotutsD. Ichihara, Y ., Jinnai, Y ., Ariu, K., Morimura, T., and Uchibe, E. Theoretical guarantees for minimum Bayes risk decod- ing. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.),Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. ...

    work page

  9. [11]

    acl-long.793/

    URL https://aclanthology.org/2025. acl-long.793/. Jinnai, Y . Document-level text generation with mini- mum Bayes risk decoding using optimal transport. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.),Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 28260–28279, Vienna, A...

    work page 2025

  10. [14]

    SVD-LLM V2: Optimizing Singular Value Truncation for Large Language Model Compression

    Association for Computational Linguistics. ISBN 979-8-89176-189-6. doi: 10.18653/v1/2025.naacl-long

    work page doi:10.18653/v1/2025.naacl-long 2025

  11. [15]

    naacl-long.472/

    URL https://aclanthology.org/2025. naacl-long.472/. Kamigaito, H., Deguchi, H., Sakai, Y ., Hayashi, K., and Watanabe, T. Diversity explains inference scaling laws: Through a case study of minimum Bayes risk decoding. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.),Proceedings of the 63rd Annual Meeting of the Association for Computationa...

    work page 2025

  12. [16]

    ISBN 979-8-89176-251-0

    Association for Computational Linguistics. ISBN 979-8-89176-251-0. doi: 10.18653/v1/2025.acl-long

    work page doi:10.18653/v1/2025.acl-long 2025

  13. [17]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018)

    URL https://aclanthology.org/2025. acl-long.1410/. Karpathy, A. and Fei-Fei, L. Deep visual-semantic align- ments for generating image descriptions. In2015 IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), pp. 3128–3137, 2015. doi: 10.1109/CVPR. 2015.7298932. URL https://ieeexplore.ieee. org/document/7298932. 12 Noisy-Channel Minimum Ba...

    work page doi:10.1109/cvpr 2025

  14. [18]

    acl-main.703/

    URL https://aclanthology.org/2020. acl-main.703/. Li, H., Zhang, M., and Su, J. A joint source-channel model for machine transliteration. InProceedings of the 42nd Annual Meeting of the Association for Compu- tational Linguistics (ACL-04), pp. 159–166, Barcelona, Spain, July 2004. doi: 10.3115/1218955.1218976. URL https://aclanthology.org/P04-1021/. Li, J...

    work page doi:10.3115/1218955.1218976 2020

  15. [19]

    ISBN 978- 3-319-10602-1

    Springer International Publishing. ISBN 978- 3-319-10602-1. doi: 10.1007/978-3-319-10602-1 48. URL https://link.springer.com/chapter/ 10.1007/978-3-319-10602-1_48. Lyu, B., Kamigaito, H., Funakoshi, K., and Okumura, M. Unveiling the power of source: Source-based minimum Bayes risk decoding for neural machine translation. In Che, W., Nabende, J., Shutova, ...

    work page doi:10.1007/978-3-319-10602-1 2025

  16. [20]

    acl-long.149/

    URL https://aclanthology.org/2025. acl-long.149/. Lyu, B., Song, H., Kamigaito, H., Ding, C., Tanaka, H., Utiyama, M., Funakoshi, K., and Okumura, M. Minimum bayes risk decoding for error span detection in reference- free automatic machine translation evaluation, 2025b. URLhttps://arxiv.org/abs/2512.07540. Mitchell, T. M.Machine Learning. McGraw-Hill, Inc...

    work page doi:10.18653/v1/2021.acl-long.22 2025

  17. [21]

    wmt-1.67/

    URL https://aclanthology.org/2023. wmt-1.67/. Natsumi, K., Deguchi, H., Sakai, Y ., Kamigaito, H., and Watanabe, T. Agreement-constrained probabilis- tic minimum Bayes risk decoding. In Inui, K., Sakti, S., Wang, H., Wong, D. F., Bhattacharyya, P., Baner- jee, B., Ekbal, A., Chakraborty, T., and Singh, D. P. (eds.),Proceedings of the 14th International Jo...

    work page 2023

  18. [22]

    Proceedings of the 40th

    URL https://aclanthology.org/2025. ijcnlp-short.39/. Och, F. J. and Ney, H. Discriminative training and maxi- mum entropy models for statistical machine translation. In Isabelle, P., Charniak, E., and Lin, D. (eds.),Proceed- ings of the 40th Annual Meeting of the Association for Computational Linguistics, pp. 295–302, Philadelphia, Pennsylvania, USA, July...

    work page doi:10.3115/1073083.1073133 2025

  19. [23]

    Proceedings of the 40th

    URL https://aclanthology.org/2024. naacl-short.38/. Ott, M., Auli, M., Grangier, D., and Ranzato, M. Ana- lyzing uncertainty in neural machine translation. In Dy, J. and Krause, A. (eds.),Proceedings of the 35th Inter- national Conference on Machine Learning, volume 80 ofProceedings of Machine Learning Research, pp. 3956–3965. PMLR, 10–15 Jul 2018. URL ht...

    work page doi:10.3115/1073083.1073135 2024

  20. [24]

    emnlp-main.213/

    URL https://aclanthology.org/2020. emnlp-main.213/. Rei, R., C. de Souza, J. G., Alves, D., Zerva, C., Farinha, A. C., Glushkova, T., Lavie, A., Coheur, L., and Martins, A. F. T. COMET-22: Unbabel-IST 2022 submission for the metrics shared task. In Koehn, P., Barrault, L., Bo- jar, O., Bougares, F., Chatterjee, R., Costa-juss`a, M. R., Federmann, C., Fish...

    work page arXiv 2020

  21. [25]

    Weber, Lena Jurkschat, Hammam Abdelwahab, Chelsea John, Pedro Ortiz Suárez, Malte Ostendorff, Samuel Weinbach, Rafet Sifa, Stefan Kesselheim, and Nicolas Flores-Herr

    URL https://openreview.net/forum? id=8iPobEKUUA. Vamvas, J. and Sennrich, R. Linear-time minimum Bayes risk decoding with reference aggregation. In Ku, L.- W., Martins, A., and Srikumar, V . (eds.),Proceedings of the 62nd Annual Meeting of the Association for Com- putational Linguistics (Volume 2: Short Papers), pp. 790–801, Bangkok, Thailand, August 2024...

    work page doi:10.18653/v1/2024 2024

  22. [26]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pages =

    Association for Computational Linguistics. doi: 10.18653/v1/2020.emnlp-demos.6. URL https:// aclanthology.org/2020.emnlp-demos.6/. Yan, J., Xu, J., Meng, F., Zhou, J., and Zhang, Y . DC- MBR: Distributional cooling for minimum Bayesian risk decoding. In Calzolari, N., Kan, M.-Y ., Hoste, V ., Lenci, A., Sakti, S., and Xue, N. (eds.),Proceedings of the 202...

    work page doi:10.18653/v1/2020.emnlp-demos.6 2020

  23. [27]

    naacl-short.34/

    URL https://aclanthology.org/2024. naacl-short.34/. Yu, L., Blunsom, P., Dyer, C., Grefenstette, E., and Kocisky, T. The neural noisy channel. InInternational Conference on Learning Representations, 2017. URL https:// openreview.net/forum?id=SJ25-B5eg. Zhang, T., Kishore, V ., Wu, F., Weinberger, K. Q., and Artzi, Y . Bertscore: Evaluating text generation...

    work page 2024

  24. [28]

    URL https://openreview.net/forum? id=SkeHuCVFDr. 15

    work page