Pith. sign in

REVIEW 4 major objections 7 minor 15 references

Document-Level Text Generation with Minimum Bayes Risk Decoding using Optimal Transport

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A document-level utility built by aligning sentences with Wasserstein distance lets MBR decoding beat its sentence-scoring baseline on translation, simplification, and dense captioning.

desk verdict A sensible, well-executed first combination of optimal-transport document distance with MBR decoding, but the 'outperforms' claim is metric-dependent and the one independent check partly reverses it. read the letter →

arxiv 2505.23078 v1 pith:WXXIQPSF submitted 2025-05-29 cs.CL cs.AI

Yuu Jinnai This is my paper
classification cs.CLcs.AI
keywords minimumBayesriskdecodingoptimaltransportWassersteindistancedocument-leveltextgenerationmachinetranslationsimplificationdenseimagecaptioningentropicregularization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Minimum Bayes Risk (MBR) decoding picks the candidate output with the highest expected utility, but the best utility functions are built for single sentences, so document-level MBR underperforms. This paper proposes MBR-OT, which splits each candidate into sentences, gives each sentence equal (or length-proportional) weight, and computes document utility as the Wasserstein distance between the candidate's and the reference's sentence-score distributions, with the cost of moving mass between two sentences equal to one minus their sentence-level utility. Because Wasserstein distance allows one sentence's mass to split across several others, merged, split, and reordered sentences are aligned fairly rather than penalized as structural mismatches. The paper reports that MBR-OT beats standard MBR on document-level machine translation, text simplification, and dense image captioning, using automated evaluation metrics, and that the translation gain also holds under an independent metric (COMET-22). If correct, this gives a decoding-time improvement that requires no retraining and can reuse the mature sentence-level utility metrics already available.

What carries the argument

The load-bearing object is the Wasserstein-distance utility of Eq. (7): the minimum cost of transporting the sentence-weight distribution of one document to that of another, using a cost matrix derived from a sentence-level utility. With the entropic regularization term of Eq. (6), it becomes EWD, which is robust to model noise and fast to compute with the Sinkhorn algorithm; this is the variant used in most experiments. Because the distance is symmetric and has zero self-distance, MBR-OT evaluates only $N(N-1)/2$ document pairs instead of $N^2$, and it can reuse state-of-the-art sentence-level metrics (MetricX-23, BERTScore, SentBERT, CLIPText) as the underlying $u_s$. The uniform or length-proportional sentence weights in Eqs. (9)-(10) determine how much each sentence's score contributes.

What would settle it

Conduct a pairwise human preference study on held-out documents from WMT24 En-Ja and En-De, JADOS, and PixelProse, comparing MBR and MBR-OT outputs with annotators blind to method; if humans do not prefer MBR-OT at rates matching or exceeding the automated metrics' margins, the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the weak point of MBR at document level is not the decoder but the utility function: sentence-level metrics cannot score a multi-sentence output as a whole, and existing document-level metrics are tied to fixed segment orders. MBR-OT defines the document-level utility of a hypothesis $h$ against a reference $y$ by $u(h, y) = 1 - \mathrm{WD}_\epsilon[p_h\|p_y]$, where $p_h$ and $p_y$ are uniform distributions over the sentences of each document and the per-sentence-pair cost is $1 - u_s(h_i, y_j)$ using a sentence-level utility $u_s$. The entropic-regularized Wasserstein distance lets probability mass be transported fractionally, so a single source sentence can be matched to several target sentences; this makes the score insensitive to reordering, merging, and splitting, which the paper argues is exactly what LLM-produced documents often contain. The paper claims that this consistently outperforms standard MBR across document-level machine translation, text simplification, and dense image captioning, with the gains robust to using a different verification metric (COMET-22) for translation.

Load-bearing premise

The paper's results rest on the assumption that the automated metrics used to compare outputs (MetricX-23-XXL with EWD for translation, D-SARI for simplification, CLAIR for dense captioning) measure what humans would regard as better document-level output; the paper itself flags that human evaluation is desirable.

Editorial extensions

If this is right

  • Document-level MBR decoding can be improved without training a document-level metric; existing sentence-level metrics become document-level utilities through optimal transport.
  • The method is robust to the variable document structures LLMs produce, since merged, split, or reordered sentences no longer depress the utility score.
  • The same algorithm applies across three different generation tasks with different underlying utilities (MetricX for translation, SentBERT for simplification and summarization, CLIPText for dense captioning), suggesting a task-independent decoding recipe.
  • Using the symmetric property of the Wasserstein distance halves the number of pairwise utility computations needed for MBR over the same candidate and reference pools.
  • The translation gains persist when evaluated with COMET-22 rather than the metric used as the utility, reducing the concern that the improvement is just self-scoring bias.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the utility ignores sentence order entirely, anaphora and other order-dependent coherence errors are invisible to it; an order-aware variant such as a Gromov-Wasserstein term is a natural next step.
  • If the advantage stems from better alignment under merges and splits rather than from metric self-scoring, the gap between MBR-OT and MBR should widen as documents become longer and more structurally variable, which a controlled length-scaling study could test.
  • A distillation-trained document-level scorer that imitates the Wasserstein utility could eliminate the decoding-time overhead the paper reports, since the paper lists distillation as an option but does not run it.
  • For open-ended generation such as story writing, where ordering and coherence carry much of the meaning, the order-blind Wasserstein utility might hurt rather than help; separating directed from open-ended tasks would clarify the boundary.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes MBR-OT, a variant of minimum Bayes risk (MBR) decoding for document-level text generation. Instead of applying a sentence-level utility function to an entire document, MBR-OT treats each document as a distribution over its sentences and defines utility as one minus the (possibly entropic regularized) Wasserstein distance between those distributions (Eq. 7). The authors evaluate the utility metric itself on WMT22/23 metric shared tasks, then test MBR-OT on WMT24 document-level machine translation (En-De, En-Ja), JADOS document simplification, CNNDM summarization, and PixelProse dense image captioning. They report that MBR-OT outperforms standard MBR on their chosen metrics and provide code, a reproducibility statement, and several appendices, including an additional COMET-22 evaluation.

Significance. The algorithmic idea is simple, general, and potentially useful: it lets a mature sentence-level utility function be applied to documents of varying structure while remaining robust to sentence reordering and merging. The paper's strengths include public code, clear notation, sensible computational observations (symmetry and low-rank approximations in Section 3), and evaluation across three diverse tasks. If the empirical claim were established with metrics independent of the selection utility and with significance testing or human evaluation, the contribution would be solid and practically relevant. As it stands, the headline is supported only on the paper's chosen metric families; the one independent metric check in Appendix A partially contradicts the MT result, and no statistical testing is reported anywhere in the manuscript.

major comments (4)
  1. [Section 4.2 / Appendix A, Tables 2 and 9] The central MT claim that MBR-OT outperforms MBR is not robust to the choice of evaluation metric. Under MetricX-23-XXL with EWD, MBR-WD(MetricX) on WMT24 En-Ja is 75.29 versus 68.81 for MBR, but under COMET-22 the same system scores 66.82 versus 67.76, i.e., it is worse than MBR; only the length-weighted MBR-WDL retains an advantage in that setting (70.46 versus 67.76). The En-De COMET-22 results also show smaller gains than in Table 2. Because the paper itself identifies the same-metric bias in Appendix A, the general 'outperforms MBR' statement in Section 6 should be conditioned on the evaluation metric and language pair, and the authors should report results with multiple independent metrics or with human evaluation.
  2. [Section 4.2] The main machine translation evaluation uses MetricX-23-XL as the selection utility and MetricX-23-XXL with EWD as the evaluation metric; both belong to the same trained metric family. This is exactly the bias acknowledged in Appendix A ('the evaluation on Section 4.2 may exhibit bias') and in the prior work cited there (Kovacs et al., 2024). The COMET-22 verification is welcome but is partial: it is reported only in an appendix, and on En-Ja it reverses the headline result for MBR-WD. The authors should present an evaluation metric fully independent of the selection utility in the main tables, or run a human evaluation as suggested in Section 7.
  3. [Sections 4.2-4.5] No significance tests, confidence intervals, or human evaluation are reported anywhere in the paper. The reported improvements are point estimates and are often modest (e.g., Table 2 En-De: MBR 82.02 versus MBR-WD 83.40; Table 8 shows no difference on METEOR for dense captioning). Appendix B's domain-level analysis uses very uneven and small samples (news=17, speech=111, literary=8, social=34). Paired bootstrap or per-item significance tests, or human evaluation, are needed to distinguish a systematic gain from sampling noise. The paper itself states in Section 7 that 'Human evaluation is desirable' for these metrics; because the central claim is an empirical superiority claim, this is a load-bearing limitation rather than a minor omission.
  4. [Sections 4.3 and 4.5] The simplification and dense image captioning evaluations each rely on a single automatic metric: D-SARI for JADOS and CLAIR for PixelProse. D-SARI is a lexical metric, and although the paper notes that lexical metrics have little correlation with human evaluation for BLEU and ROUGE, no alternative semantic metric or human judgment is provided for the simplification outputs. For dense captioning, METEOR shows no difference between algorithms in Table 8, so the CLAIR improvements alone are not sufficient to establish that MBR-OT produces higher-quality captions. The authors should add at least one independent metric or a qualitative/human analysis for these tasks before claiming consistent superiority.
minor comments (7)
  1. [Section 3, Eq. (12) and surrounding text] The sentence 'the value of WD is 1 when the two distributions are the same' is incorrect for a Wasserstein distance metric, which is 0 for identical inputs; consequently the displayed 'u(h, h) = 0' is also inconsistent with the definition u = 1 - WD in Eq. (7). The text likely intends 'WD is 0' and 'u(h, h) = 1'.
  2. [Section 1] The sentence 'Our results show that MBR decoding consistently outperforms the baselines' appears to be a typo: the intended claim, as stated in the abstract and Section 6, is that MBR-OT outperforms the baselines including MBR decoding.
  3. [Section 3] The notation for the linear assignment variant is inconsistent: the text introduces 'MBR-AL' but the experimental tables and figures use 'MBR-LA'; please unify the notation.
  4. [Section 2.2, Eq. (10)] The displayed formula for the length-proportional weight is rendered as '|hi|P j |hj|' without a division sign; it should read '|hi| / sum_j |hj|'.
  5. [Section 4.2 / Table 4] The sentence 'the performance of MBR-OT is outperforming the baselines in both language pairs' is too strong when the BLEU reference in Table 4 shows MBR-WD-epsilon on En-Ja (9.66) slightly below MBR (10.01). Even if BLEU is only a reference metric, the claim should be qualified.
  6. [Appendix B] The domain-level analysis in Table 11 should be presented with a clear caveat in the main text because the per-domain document counts are very small, especially for news (17) and literary (8), making the domain-level comparisons statistically fragile.
  7. [Various] Minor typographical issues: 'sophisticted' in Section 7, 'spaPy' in Section 4 for spaCy, and 'propotional' in Section 3; these should be corrected in revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MBR-OT is a direct extension of MBR decoding (Eq. 7), the claimed gains are empirical, and the acknowledged metric-bias caveats are validity concerns rather than circular reductions.

full rationale

The central derivation is self-contained. Equation (7) defines the document-level utility as u(h, y) = 1 - OT[p_h || p_y], with the per-sentence cost C(h_i, y_j) = 1 - u_s(h_i, y_j) in Eq. (8). Plugging this utility into the standard MBR objective of Eq. (3) is a definitional extension, not an inference from the data that later appears as the conclusion. No parameter is fitted to the headline 'outperforms' claim, and the experimental comparisons are against external benchmarks and baselines. The main caveat is evaluation-metric overlap: in Section 4.2, selection uses MetricX-23-XL while the main evaluation uses MetricX-23-XXL with EWD. The paper itself flags this in Appendix A ('Using the same evaluation metric as the utility function that MBR decoding uses is known to cause a bias in the evaluation... the evaluation on Section 4.2 may exhibit bias') and supplies COMET-22 as an independent verification. That independent check is mixed rather than uniformly supportive, and Section 7 notes that 'The study depends on automated metrics... Human evaluation is desirable.' These are legitimate empirical-validity limitations, but they do not make any equation or selection rule reduce by construction to its inputs. The self-citations to Jinnai et al. (2024) for epsilon sampling and to Jinnai and Ariu (2024) for faster MBR are not load-bearing for the MBR-OT result; epsilon sampling is also credited to the independent Freitag et al. (2023a), and those sampling choices do not define the proposed Wasserstein utility. No circular step can be exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The method introduces no new entities or fitted constants beyond epsilon and the weighting choice. The main burden is not in the mathematics but in the assumptions that sentence-level metrics and automated document metrics capture the quality that humans would assign.

free parameters (2)
  • EWD regularization coefficient epsilon = 0.1 (WD uses 0)
    Chosen in Section 4.2 without a tuning study; determines the entropic regularization in Eq. 6 and is used for the main results in Tables 2, 3, 9, and 10.
  • Segment weighting scheme = uniform (Eq. 9) or length-proportional (Eq. 10)
    A hand-selected modeling choice. Results depend on it: MBR-WDL beats MBR-WD on COMET-22 En-Ja in Appendix A, while MBR-WD loses to MBR there.
assumptions (6)
  • domain assumption The Monte Carlo estimate in Eq. (3) is a faithful approximation of the expected utility in Eq. (1) for long documents.
    Standard MBR practice, but sample coverage of the posterior is not guaranteed for long documents with many structural variations.
  • standard math Optimal transport definitions and the Sinkhorn algorithm are taken as standard background (Eqs. 4-6).
    No proof is given beyond citations to Peyré-Cuturi and Villani; this is a reasonable background assumption.
  • domain assumption Sentence boundaries and uniform or length weights summarize the information structure of a document.
    The method's utility depends on this segmentation and weighting; the paper notes in Section 7 that more sophisticated weighting is future work.
  • domain assumption Automated metrics used for evaluation (MetricX-23 under EWD, D-SARI, CLAIR) correlate with human judgment for documents.
    Section 7 states: "The study depends on automated metrics... Human evaluation is desirable." This is load-bearing for the empirical claim.
  • domain assumption MetricX-23 with EWD is a valid evaluation metric because it shows high correlation in Table 1.
    Section 4.2 explicitly justifies the evaluation this way, but Table 1 contains anomalies such as the WMT23 he-en correlations dropping sharply, so the inference is fragile.
  • domain assumption Ordering of sentences beyond segmentation is non-essential for utility; coreference and order effects are ignorable.
    Section 7 admits that pronoun meaning depends on sentence order and the OT formulations cannot distinguish such cases.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Document-Level Text Generation with Minimum Bayes Risk Decoding using Optimal Transport." pith.science (2026). https://pith.science/paper/WXXIQPSF

@misc{pith2026250523078,
  author       = {Pith},
  title        = {Pith review of: Document-Level Text Generation with Minimum Bayes Risk Decoding using Optimal Transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WXXIQPSF}},
  note         = {Machine review of arXiv:2505.23078}
}
read the original abstract

Document-level text generation tasks are known to be more difficult than sentence-level text generation tasks as they require the understanding of longer context to generate high-quality texts. In this paper, we investigate the adaption of Minimum Bayes Risk (MBR) decoding for document-level text generation tasks. MBR decoding makes use of a utility function to estimate the output with the highest expected utility from a set of candidate outputs. Although MBR decoding is shown to be effective in a wide range of sentence-level text generation tasks, its performance on document-level text generation tasks is limited as many of the utility functions are designed for evaluating the utility of sentences. To this end, we propose MBR-OT, a variant of MBR decoding using Wasserstein distance to compute the utility of a document using a sentence-level utility function. The experimental result shows that the performance of MBR-OT outperforms that of the standard MBR in document-level machine translation, text simplification, and dense image captioning tasks. Our code is available at https://github.com/jinnaiyuu/mbr-optimal-transport

Figures

Figures reproduced from arXiv: 2505.23078 by the authors.

Figure 1
Figure 1. Illustrative example of a metric using Wasser [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Evaluation of MBR-OT on document-level machine translation tasks. WD metric with MetricX￾23 as a sentence-level utility function is used as the evaluation metric. (google/metricx-23-xxl-v2p0) to alleviate the overfitting problem of MBR decoding (Kovacs et al., 2024). We additionally evaluate with COMET-22 (Unbabel/wmt22-comet-da; Rei et al. 2022) in Appendix A. Comparison of LA, WD, and EWD [PITH_FULL_IMAGE:figures… view at source ↗
Figure 3
Figure 3. Evaluation of MBR-OT on document-level summarization and simplification tasks. metric to evaluate the outputs. CLAIR uses GPT-4 as a judge to compute the utility of the caption [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Evaluation of MBR-OT on dense image cap [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Evaluation of MBR-OT on document-level machine translation tasks using COMET-22. Llama-3.1 is used as the text generation model. third example, the first sentence of the document appears to be interpreted as the title of the article. Although the document-level MT is a…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 10 canonical work pages

  1. [4]

    BlonDe: An automatic evaluation metric for document-level machine translation. In Proceedings of the 2022 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies, pages 1550–1565, Seattle, United States. Association for Computational Linguistics. Yuu Jinnai and Kaito Ariu. 2024. Hyperparamet...

    work page 2022

  2. [5]

    In Proceedings of the Eighth Conference on Machine Translation , pages 756–767, Singapore

    MetricX-23: The Google submission to the WMT 2023 metrics shared task. In Proceedings of the Eighth Conference on Machine Translation , pages 756–767, Singapore. Association for Compu- tational Linguistics. Tom Kocmi, Eleftherios Avramidis, Rachel Bawden, Ondˇrej Bojar, Anton Dvorkovich, Christian Feder- mann, Mark Fishel, Markus Freitag, Thamme Gowda, Ro...

    work page 2023

  3. [6]

    In Proceedings of the Ninth Conference on Machine Translation, pages 1063–1094, Miami, Florida, USA

    Mitigating metric bias in minimum Bayes risk decoding. In Proceedings of the Ninth Conference on Machine Translation, pages 1063–1094, Miami, Florida, USA. Association for Computational Lin- guistics. Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin John- son, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al

    work page

  4. [9]

    In Proceedings of the 61st An- nual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 223–241, Toronto, Canada

    Open-ended long text generation via masked language modeling. In Proceedings of the 61st An- nual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 223–241, Toronto, Canada. Association for Computational Lin- guistics. Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and L...

    work page 2020

  5. [10]

    In Proceedings of the 2024 Joint In- ternational Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING 2024), pages 459–476, Torino, Italia

    A document-level text simplification dataset for Japanese. In Proceedings of the 2024 Joint In- ternational Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING 2024), pages 459–476, Torino, Italia. ELRA and ICCL. Yasumasa Onoe, Sunayana Rane, Zachary Berger, Yonatan Bitton, Jaemin Cho, Roopal Garg, Alexan- der Ku, Za...

    arXiv 2024

  6. [11]

    In Thirty-seventh Conference on Neural Information Processing Sys- tems

    Direct preference optimization: Your language model is secretly a reward model. In Thirty-seventh Conference on Neural Information Processing Sys- tems. Miguel Ramos, Patrick Fernandes, António Farinhas, and Andre Martins. 2024. Aligning neural machine translation models: Human feedback in training and inference. In Proceedings of the 25th Annual Con- fer...

    work page 2024

  7. [12]

    In Proceedings of the Seventh Conference on Machine Translation (WMT) , pages 578–585, Abu Dhabi, United Arab Emirates (Hybrid)

    COMET-22: Unbabel-IST 2022 submission for the metrics shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT) , pages 578–585, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics. Vasu Singla, Kaiyu Yue, Sukriti Paul, Reza Shirka- vand, Mayuka Jayawardhana, Alireza Ganjdanesh, Heng Huang, Abhinav ...

    arXiv 2022

  8. [14]

    People Swim- ming in the Swimming Pool

    and are less structured than those in news domains sourced from online news sites. OT may COMET-22 En-Ja En-De Beam 58.14 67.19 MBR (SFR2) 53.98 61.72 MBR (MetricX) 67.76 70.77 MBR-LA (MetricX) 64.12 72.56 MBR-LAL (MetricX) 64.56 73.17 MBR-WD (MetricX) 66.82 72.70 MBR-WDL (MetricX) 70.46 73.13 MBR-WDϵ (MetricX) 66.82 73.62 MBR-WDϵ L (MetricX) 70.23 73.86 ...

    work page 2022

Show all 15 references
  1. [15]

    for all the experiments as it is shown to be effective for generating samples for MBR decoding (Freitag et al., 2023a; Jinnai et al., 2024). 18 Parameter Value Temperature 1.0 top_p 1.0 epsilon_cutoff 0.01 max_new_tokens 1024 Table 12: Hyperparmeters for text generation Parame...

    2022

  2. [2004]

    In Proceedings of the 2004 Conference on Empirical Methods in Natu- ral Language Processing, pages 230–237, Barcelona, Spain

    Applying conditional random fields to Japanese morphological analysis. In Proceedings of the 2004 Conference on Empirical Methods in Natu- ral Language Processing, pages 230–237, Barcelona, Spain. Association for Computational Linguistics. 12 Shankar Kumar and William Byrne. 2...

    2004

  3. [2017]

    International journal of computer vision , 123:32–73

    Visual genome: Connecting language and vi- sion using crowdsourced dense image annotations. International journal of computer vision , 123:32–73. Keito Kudo, Hiroyuki Deguchi, Makoto Morishita, Ryo Fujii, Takumi Ito, Shintaro Ozaki, Koki Natsumi, Kai Sato, Kazuki Yano, Ryosuke...

    2024

  4. [2022]

    In Findings of the Association for Com- putational Linguistics: EMNLP 2022 , pages 3414– 3427, Abu Dhabi, United Arab Emirates

    Truncation sampling as language model desmoothing. In Findings of the Association for Com- putational Linguistics: EMNLP 2022 , pages 3414– 3427, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. John T. Holodnak and Ilse C. F. Ipsen. 2015. Ran- domiz...

    2022

  5. [2023]

    ESAIM: Control, Optimisation and Calculus of V ariations, 29:33

    Regularization for wasserstein distributionally robust optimization. ESAIM: Control, Optimisation and Calculus of V ariations, 29:33. Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with im- proved correlation with human judgments. In Pro...

    2005 arXiv

  6. [2024]

    Transac- tions of the Association for Computational Linguis- tics, 12:979–995

    xcomet: Transparent machine translation eval- uation through fine-grained error detection. Transac- tions of the Association for Computational Linguis- tics, 12:979–995. Kamil Guttmann, Mikołaj Pokrywka, Adrian Charkiewicz, and Artur Nowakowski. 2024. Chasing COMET: Leveraging...

    2024 arXiv

  7. [6284]

    PMLR. Hugo Touvron, Louis Martin, Kevin Stone, Peter Al- bert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jere...

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.