Pith. sign in

REVIEW 2 minor 42 references

Non-autoregressive models improve speech recognition accuracy by selecting outputs that maximize expected utility over samples drawn in one forward pass.

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 · grok-4.3

2026-06-26 23:11 UTC pith:FONGVOS6

load-bearing objection NAR-MBR gives a practical way to boost accuracy in fast speech recognition models by using cheap parallel samples for MBR.

arxiv 2606.17537 v1 pith:FONGVOS6 submitted 2026-06-16 eess.AS cs.CL

Non-Autoregressive Minimum Bayes' Risk Decoding for Fast Speech Recognition

Hiroyuki Deguchi , Takatomo Kano , Katsuki Chousa , Marc Delcroix This is my paper
classification eess.AS cs.CL
keywords non-autoregressive decodingminimum Bayes riskspeech recognitionfast inferenceLibriSpeechSwitchboardAMI
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 paper proposes NAR-MBR decoding to close the accuracy gap between non-autoregressive and autoregressive speech recognition. Standard NAR decoding generates all tokens in parallel and therefore faster, but cannot resolve token uncertainty by conditioning on earlier choices. NAR-MBR instead draws multiple candidate outputs from the NAR probability distribution in a single forward computation, then chooses the hypothesis that minimizes expected risk under a utility function. Experiments on LibriSpeech, Switchboard, AMI, and a web presentation corpus show higher accuracy than prior NAR methods and lower latency than autoregressive decoding. A reader would care because the approach preserves the core speed advantage of parallel generation while adding a cheap risk-aware selection step.

Core claim

NAR-MBR decoding maximizes the expected utility calculated from samples drawn from the output probability of an NAR model rather than maximizing the output probability itself. Because NAR models produce the full distribution in one forward pass, multiple independent samples can be obtained without repeated model evaluations, allowing an efficient approximation of minimum Bayes risk decoding.

What carries the argument

NAR-MBR decoding, which approximates minimum Bayes risk by drawing multiple output samples from a non-autoregressive model's distribution in a single forward computation.

Load-bearing premise

Multiple independent samples can be drawn from the NAR model's output distribution with only a single forward pass.

What would settle it

An experiment in which NAR-MBR decoding produces no accuracy gain over standard NAR beam search on the same corpora, or in which its wall-clock time exceeds that of autoregressive decoding.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • NAR-MBR decoding outperforms previous NAR decoding methods on LibriSpeech, Switchboard, AMI, and web presentation data.
  • NAR-MBR decoding runs faster than autoregressive decoding while improving accuracy.
  • The single-forward-pass sampling property enables cheap approximation of expected utility without extra model evaluations.
  • The framework directly addresses the uncertainty-resolution limitation of non-autoregressive generation.

Where Pith is reading between the lines

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

  • The same single-pass sampling trick could be tested on non-autoregressive models for machine translation or text generation if their output distributions allow independent draws.
  • If the utility function can be chosen to match downstream task metrics, NAR-MBR might reduce the need for task-specific fine-tuning of the base NAR model.
  • The approach implies that much of the NAR performance gap stems from selection rather than from the inability to model left-to-right dependencies.

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

0 major / 2 minor

Summary. The manuscript proposes NAR-MBR decoding, a non-autoregressive framework for speech recognition that applies minimum Bayes' risk decoding by maximizing expected utility over samples drawn from the NAR model's per-position output distributions. Multiple samples are obtained at negligible cost via repeated multinomial sampling after a single forward pass. Experiments across LibriSpeech, Switchboard, AMI, and a web presentation corpus are reported to show that NAR-MBR outperforms prior NAR methods while remaining faster than autoregressive decoding.

Significance. If the experimental results hold, the work demonstrates a practical route to close the accuracy gap between NAR and AR decoding in ASR without sacrificing the parallel-generation speed advantage. The single-forward-pass sampling mechanism is a direct and efficient exploitation of the NAR architecture and constitutes a clear technical contribution.

minor comments (2)
  1. [Abstract] Abstract: the claim of outperformance on four corpora is stated without any numerical results, baselines, or error bars, which weakens the ability of readers to immediately gauge the magnitude and reliability of the reported gains.
  2. [Method] The description of the MBR approximation would benefit from an explicit statement of the utility function employed and the number of samples used in the expectation estimate.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the careful reading and positive summary of our work on NAR-MBR decoding. The recommendation for minor revision is noted. As no specific major comments were raised, we provide no point-by-point responses below.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper presents NAR-MBR as a direct application of existing MBR decoding to NAR models, with the efficiency gain following immediately from the parallel sampling property of NAR (one forward pass yields per-position distributions for repeated multinomial draws). No equations, ansatzes, or fitted parameters are defined in terms of the target quantity; the central performance claims rest on external experimental results across four corpora rather than any self-referential reduction. No self-citation is used as a load-bearing uniqueness theorem or to smuggle in an ansatz. The derivation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no free parameters, axioms, or invented entities are identifiable from the provided text.

pith-pipeline@v0.9.1-grok · 5678 in / 1025 out tokens · 32161 ms · 2026-06-26T23:11:50.353047+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Non-Autoregressive Minimum Bayes' Risk Decoding for Fast Speech Recognition." pith.science (2026). https://pith.science/paper/FONGVOS6

@misc{pith2026260617537,
  author       = {Pith},
  title        = {Pith review of: Non-Autoregressive Minimum Bayes' Risk Decoding for Fast Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FONGVOS6}},
  note         = {Machine review of arXiv:2606.17537}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Non-autoregressive (NAR) decoding generates output tokens in parallel, making speech recognition faster than autoregressive decoding, which generates them sequentially from left to right. However, the recognition performance is degraded because NAR decoding cannot resolve uncertainty by conditioning on previously generated tokens. To address this issue, we propose a novel NAR decoding framework based on minimum Bayes' risk (MBR) decoding, termed NAR-MBR decoding, that maximizes the expected utility calculated from samples drawn from the output probability of an NAR model rather than maximizing the output probability. Notably, by leveraging the nature of NAR models, multiple samples are obtained efficiently with a single forward computation. Our experiments across LibriSpeech, Switchboard, AMI, and web presentation corpus demonstrated that our NAR-MBR decoding outperformed previous NAR decoding and ran faster than AR decoding.

Figures

Figures reproduced from arXiv: 2606.17537 by Hiroyuki Deguchi, Katsuki Chousa, Marc Delcroix, Takatomo Kano.

Figure 1
Figure 1. Figure 1: NAR and NAR-MBR decoding in Mask-CTC at Niter = 1. Differences from NAR decoding are highlighted . Non-autoregressive decoding. For faster decoding, NAR de￾coding reduces the number of probability calculations. It gen￾erates multiple tokens in parallel, based on the independence assumption between output tokens, as follows: pNAR(y|x; θ) := Q|y| t=1 p(yt|x). (2) Mask-CTC [8], one of the most popular non-aut… view at source ↗
Figure 2
Figure 2. Figure 2: WER of NAR-MBR decoding at Niter = 1 when vary￾ing number of samples in dev. set (Other) of LibriSpeech ing with 64 samples at Niter = 1 was faster, yet it significantly improved the WER. The memory usage of NAR-MBR decod￾ing was comparable to that of other methods when Niter = 0 but increased when Niter = 1, primarily due to the computational cost of the CMLM decoder. Addressing this issue through more me… 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

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

  1. [1]

    Introduction Decoding speed in automatic speech recognition (ASR) is chal- lenging in real-world scenarios, especially when processing long speech signals. Most of the recent state-of-the-art ASR models that have achieved low word error rates (WER) gener- ate transcriptions in a left-to-right manner, called autoregressive (AR) decoding, which takes comput...

  2. [2]

    Non-Autoregressive Minimum Bayes' Risk Decoding for Fast Speech Recognition

    Background: Decoding in ASR LetXbe the input space of ASR models andY :=V ∗ be the output space, whereV ∗ is the Kleene closure of vocabularyV. The goal of ASR is to generate a token sequencey∈ Yfrom an input speechx∈ X. Output tokensyare converted into the transcription textσ(y)∈Σ ∗, whereΣ ∗ is the Kleene closure of alphabetΣ, andσ:V ∗ →Σ ∗ denotes a de...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    Figure 1(b) presents the overview of NAR-MBR decod- ing

    Proposed Method: NAR-MBR Decoding Our non-autoregressive minimum Bayes’ risk (NAR-MBR) de- coding aims to decode high-quality transcriptions significantly faster than AR decoding, without additional training for Mask- CTC. Figure 1(b) presents the overview of NAR-MBR decod- ing. It mainly consists of two steps: (1) probabilistic unbiased sampling from Mas...

  4. [4]

    Clean” and “Other

    Experiments 4.1. Setup To confirm the effectiveness of NAR-MBR decoding, we evalu- ate recognition performance and decoding speed in ASR tasks. Datasets.We use LibriSpeech [20], Switchboard (SWBD) [21], AMI [22], and web presentation corpus (Web). The Web cor- pus consists of 346 hours of training data from 1,938 speakers, and 3.7 hours of development and...

  5. [5]

    The proposed method addresses the performance degradation of NAR decoding via probabilistic sampling and EU maximization

    Conclusion This paper proposes NAR-MBR decoding for fast speech recognition. The proposed method addresses the performance degradation of NAR decoding via probabilistic sampling and EU maximization. Experimental results demonstrated that our NAR-MBR decoding outperforms previous NAR decoding without additional training, and achieves a better trade-off be-...

  6. [6]

    All implementations, experi- ments, and analyses were conducted by the authors, who take full responsibility for the content

    Generative AI use disclosure Our use of generative AI tools was limited to proofreading, such as grammar and spell-checkers. All implementations, experi- ments, and analyses were conducted by the authors, who take full responsibility for the content

  7. [7]

    Conformer: Convolution-augmented Transformer for Speech Recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented Transformer for Speech Recognition,” in Interspeech 2020, 2020, pp. 5036–5040

    2020

  8. [8]

    E-branchformer: Branchformer with enhanced merging for speech recognition,

    K. Kim, F. Wu, Y . Peng, J. Pan, P. Sridhar, K. J. Han, and S. Watanabe, “E-branchformer: Branchformer with enhanced merging for speech recognition,” in2022 IEEE Spoken Language Technology Workshop (SLT), 2023, pp. 84–91

    2023

  9. [9]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. Mcleavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inProceedings of the 40th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett, Eds., vol. 202. PMLR, 23–2...

    2023

  10. [10]

    Less is More: Accu- rate Speech Recognition & Translation without Web-Scale Data,

    K. C. Puvvada, P. ˙Zelasko, H. Huang, O. Hrinchuk, N. R. Koluguri, K. Dhawan, S. Majumdar, E. Rastorgueva, Z. Chen, V . Lavrukhin, J. Balam, and B. Ginsburg, “Less is More: Accu- rate Speech Recognition & Translation without Web-Scale Data,” inInterspeech 2024, 2024, pp. 3964–3968

    2024

  11. [11]

    Non-autoregressive neural machine translation,

    J. Gu, J. Bradbury, C. Xiong, V . O. Li, and R. Socher, “Non-autoregressive neural machine translation,” inInternational Conference on Learning Representations, 2018

    2018

  12. [12]

    End-to-end non-autoregressive neural machine translation with connectionist temporal classification,

    J. Libovický and J. Helcl, “End-to-end non-autoregressive neural machine translation with connectionist temporal classification,” inProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, E. Riloff, D. Chiang, J. Hockenmaier, and J. Tsujii, Eds. Brussels, Belgium: Association for Computational Linguistics, Oct.-Nov. 2018, p...

    2018

  13. [13]

    Mask-predict: Parallel decoding of conditional masked lan- guage models,

    M. Ghazvininejad, O. Levy, Y . Liu, and L. Zettlemoyer, “Mask-predict: Parallel decoding of conditional masked lan- guage models,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), K. Inui, J. Jiang, V . Ng, and X. Wan, Eds. Hon...

    2019

  14. [14]

    Mask CTC: Non-Autoregressive End-to-End ASR with CTC and Mask Predict,

    Y . Higuchi, S. Watanabe, N. Chen, T. Ogawa, and T. Kobayashi, “Mask CTC: Non-Autoregressive End-to-End ASR with CTC and Mask Predict,” inInterspeech 2020, 2020, pp. 3655–3659

    2020

  15. [15]

    von Neumann and O

    J. von Neumann and O. Morgenstern,Theory of Games and Eco- nomic Behavior. Princeton: Princeton University Press, 1944

    1944

  16. [16]

    Discriminative training for segmental minimum bayes risk decoding,

    V . Doumpiotis, S. Tsakalidis, and W. Byrne, “Discriminative training for segmental minimum bayes risk decoding,” in2003 IEEE International Conference on Acoustics, Speech, and Signal Processing, 2003. Proceedings. (ICASSP ’03)., vol. 1, 2003, pp. 136–139

    2003

  17. [17]

    Lattice segmen- tation and minimum Bayes risk discriminative training,

    V . Doumpiotis, S. Tsakalidis, and W. J. Byrne, “Lattice segmen- tation and minimum Bayes risk discriminative training,” in8th European Conference on Speech Communication and Technology (Eurospeech 2003), 2003, pp. 1985–1988

    2003

  18. [18]

    Minimum word error rate training for attention-based sequence-to-sequence models,

    R. Prabhavalkar, T. N. Sainath, Y . Wu, P. Nguyen, Z. Chen, C.- C. Chiu, and A. Kannan, “Minimum word error rate training for attention-based sequence-to-sequence models,” in2018 IEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 2018, pp. 4839–4843

    2018

  19. [19]

    Minimum bayes risk training for end-to-end speaker-attributed asr,

    N. Kanda, Z. Meng, L. Lu, Y . Gaur, X. Wang, Z. Chen, and T. Yoshioka, “Minimum bayes risk training for end-to-end speaker-attributed asr,” inICASSP 2021 - 2021 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 6503–6507

    2021

  20. [20]

    Bayes risk ctc: Controllable ctc alignment in sequence-to-sequence tasks,

    J. Tian, B. Yan, J. Yu, C. Weng, D. Yu, and S. Watanabe, “Bayes risk ctc: Controllable ctc alignment in sequence-to-sequence tasks,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  21. [21]

    Minimum bayes-risk automatic speech recognition,

    V . Goel and W. J. Byrne, “Minimum bayes-risk automatic speech recognition,”Computer Speech & Language, vol. 14, no. 2, pp. 115–135, 2000

    2000

  22. [22]

    Minimum Bayes-risk decoding for statistical machine translation,

    S. Kumar and W. Byrne, “Minimum Bayes-risk decoding for statistical machine translation,” inProceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT- NAACL 2004. Boston, Massachusetts, USA: Association for Computational Linguistics, May 2 - May 7 2004, pp. 169–176

    2004

  23. [23]

    Task-specific minimum Bayes-risk de- coding using learned edit distance,

    I. Shafran and W. Byrne, “Task-specific minimum Bayes-risk de- coding using learned edit distance,” inInterspeech 2004, 2004, pp. 1945–1948

    2004

  24. [24]

    An improved consensus- like method for minimum bayes risk decoding and lattice com- bination,

    H. Xu, D. Povey, L. Mangu, and J. Zhu, “An improved consensus- like method for minimum bayes risk decoding and lattice com- bination,” in2010 IEEE International Conference on Acoustics, Speech and Signal Processing, 2010, pp. 4938–4941

    2010

  25. [25]

    Is MAP decoding all you need? the inadequacy of the mode in neural machine translation,

    B. Eikema and W. Aziz, “Is MAP decoding all you need? the inadequacy of the mode in neural machine translation,” inProceedings of the 28th International Conference on Computational Linguistics, D. Scott, N. Bel, and C. Zong, Eds. Barcelona, Spain (Online): International Committee on Computational Linguistics, Dec. 2020, pp. 4506–4520

    2020

  26. [26]

    Lib- rispeech: An asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: An asr corpus based on public domain audio books,” in2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206–5210

    2015

  27. [27]

    Switchboard: telephone speech corpus for research and development,

    J. J. Godfrey, E. C. Holliman, and J. McDaniel, “Switchboard: telephone speech corpus for research and development,” inPro- ceedings of the 1992 IEEE International Conference on Acous- tics, Speech and Signal Processing - Volume 1, ser. ICASSP’92. USA: IEEE Computer Society, 1992, p. 517–520

    1992

  28. [28]

    The ami meeting corpus,

    I. McCowan, J. Carletta, W. Kraaij, S. Ashby, S. Bourban, M. Flynn, M. Guillemot, T. Hain, J. Kadlec, and V . Karaiskos, “The ami meeting corpus,” inInternational Conference on Meth- ods and Techniques. UIA, 2005

    2005

  29. [29]

    Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks,

    A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, “Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks,” inProceedings of the 23rd International Conference on Machine Learning, ser. ICML ’06. New York, NY , USA: Association for Computing Machinery, 2006, p. 369–376

    2006

  30. [30]

    Six challenges for neural machine translation,

    P. Koehn and R. Knowles, “Six challenges for neural machine translation,” inProceedings of the First Workshop on Neural Machine Translation, T. Luong, A. Birch, G. Neubig, and A. Finch, Eds. Vancouver: Association for Computational Linguistics, Aug. 2017, pp. 28–39

    2017

  31. [31]

    Analyzing uncer- tainty in neural machine translation,

    M. Ott, M. Auli, D. Grangier, and M. Ranzato, “Analyzing uncer- tainty in neural machine translation,” inInternational Conference on Machine Learning, 2018

    2018

  32. [32]

    Sampling-based approximations to minimum Bayes risk decoding for neural machine translation,

    B. Eikema and W. Aziz, “Sampling-based approximations to minimum Bayes risk decoding for neural machine translation,” inProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Y . Goldberg, Z. Kozareva, and Y . Zhang, Eds. Abu Dhabi, United Arab Emirates: Association for Computational Linguistics, Dec. 2022, pp. 10 978–10 993

    2022

  33. [33]

    Epsilon sampling rocks: Investigating sampling strategies for minimum Bayes risk decoding for machine translation,

    M. Freitag, B. Ghorbani, and P. Fernandes, “Epsilon sampling rocks: Investigating sampling strategies for minimum Bayes risk decoding for machine translation,” inFindings of the Association for Computational Linguistics: EMNLP 2023, H. Bouamor, J. Pino, and K. Bali, Eds. Singapore: Association for Computational Linguistics, Dec. 2023, pp. 9198–9209

    2023

  34. [34]

    Gumbel,Statistical Theory of Extreme Values and Some Practical Applications: A Series of Lectures, ser

    E. Gumbel,Statistical Theory of Extreme Values and Some Practical Applications: A Series of Lectures, ser. Applied mathematics series. U.S. Government Printing Office, 1954

    1954

  35. [35]

    ESPnet: End-to-End Speech Processing Toolkit,

    S. Watanabe, T. Hori, S. Karita, T. Hayashi, J. Nishitoba, Y . Unno, N. Enrique Yalta Soplin, J. Heymann, M. Wiesner, N. Chen, A. Renduchintala, and T. Ochiai, “ESPnet: End-to-End Speech Processing Toolkit,” inInterspeech 2018, 2018, pp. 2207–2211

    2018

  36. [36]

    Joint ctc-attention based end-to-end speech recognition using multi-task learning,

    S. Kim, T. Hori, and S. Watanabe, “Joint ctc-attention based end-to-end speech recognition using multi-task learning,” in2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 4835–4839

    2017

  37. [37]

    Joint CTC/attention decoding for end-to-end speech recognition,

    T. Hori, S. Watanabe, and J. Hershey, “Joint CTC/attention decoding for end-to-end speech recognition,” inProceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), R. Barzilay and M.-Y . Kan, Eds. Vancouver, Canada: Association for Computational Linguistics, Jul. 2017, pp. 518–529

    2017

  38. [38]

    Statistical significance tests for machine translation evaluation,

    P. Koehn, “Statistical significance tests for machine translation evaluation,” inProceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, D. Lin and D. Wu, Eds. Barcelona, Spain: Association for Computational Linguistics, Jul. 2004, pp. 388–395

    2004

  39. [39]

    Bootstrap estimates for confidence inter- vals in asr performance evaluation,

    M. Bisani and H. Ney, “Bootstrap estimates for confidence inter- vals in asr performance evaluation,” in2004 IEEE International Conference on Acoustics, Speech, and Signal Processing, vol. 1, 2004, pp. 409–412

    2004

  40. [40]

    mbrs: A library for minimum Bayes risk decoding,

    H. Deguchi, Y . Sakai, H. Kamigaito, and T. Watanabe, “mbrs: A library for minimum Bayes risk decoding,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, D. I. Hernandez Farias, T. Hope, and M. Li, Eds. Miami, Florida, USA: Association for Computational Linguistics, Nov. 2024, pp. 351–362

    2024

  41. [41]

    Theoretical guarantees for minimum Bayes risk decoding,

    Y . Ichihara, Y . Jinnai, K. Ariu, T. Morimura, and E. Uchibe, “Theoretical guarantees for minimum Bayes risk decoding,” in Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar, Eds. Vienna, Austria: Association for Computational Linguistics, Jul...

    2025

  42. [42]

    Diversity explains inference scaling laws: Through a case study of minimum Bayes risk decoding,

    H. Kamigaito, H. Deguchi, Y . Sakai, K. Hayashi, and T. Watanabe, “Diversity explains inference scaling laws: Through a case study of minimum Bayes risk decoding,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar, Eds. Vienna, Austria: Asso...

    2025