Pith. sign in

REVIEW 5 major objections 4 minor 2 cited by

A simultaneous speech translator can be trained from sentence-level pairs alone, with reinforcement learning replacing the word-aligned data bottleneck.

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 · deepseek-v4-flash

2026-08-03 00:52 UTC pith:BNPUQQLW

load-bearing objection Useful method that removes word-level alignment for simultaneous S2ST, but the SOTA claim leans on a self-built synthetic benchmark and omits two strong cited baselines. the 5 major comments →

arxiv 2602.11072 v2 pith:BNPUQQLW submitted 2026-02-11 cs.CL cs.SDeess.AS

Simultaneous Speech-to-Speech Translation Without Aligned Data

Tom Labiausse , Romain Fabre , Yannick Est\`eve , Alexandre D\'efossez , Neil Zeghidour This is my paper
classification cs.CL cs.SDeess.AS
keywords simultaneous speech translationspeech-to-speech translationreinforcement learningprocess rewardsBLEUmultistream audio modelinglow-latency translationzero-alignment training
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.

This paper tries to establish that simultaneous speech-to-speech translation does not need word-level aligned training data, which has long been the bottleneck for scaling this technology. Instead, it trains a base model on sentence-level aligned pairs and then uses reinforcement learning with a simple BLEU-based reward to teach the model when to start speaking. The authors claim this approach matches or beats previous supervised systems on translation quality, latency, speaker identity preservation, and naturalness across French, Spanish, Portuguese, German, and a new-language adaptation setting. If true, adding a language reduces to collecting sentence-aligned audio, not designing language-specific alignment heuristics.

Core claim

Hibiki-Zero is a decoder-only multistream model that maps source speech tokens directly to target speech tokens through an inner text stream. The paper's central claim is that word-level alignment -- the usual prerequisite for training simultaneous translation -- is unnecessary. The authors first train a base model on sentence-level aligned pairs, inserting artificial silences into the target speech to create coarse temporal overlap, then refine it with reinforcement learning. The reward is BLEU computed at intermediate frames against the known translation of the source sentences heard so far, normalized across a group of sampled translations and accumulated into advantages. This process rew

What carries the argument

The load-bearing device is the BLEU-based process reward computed over the model's own inner-monologue text stream at the end of every nw input words. For each sampled translation, the model's partial text up to a frame is compared with the reference translation of the source sentences heard so far; the reward is a weighted blend of that partial BLEU and the final BLEU. Rewards are normalized across the group of samples at each frame and accumulated as advantages into a clipped GRPO objective over all output codebooks. The multistream RQ-Transformer -- a residual-quantized audio codec modeled along a time axis and a codebook-depth axis -- is what allows the same network to listen, generate a

Load-bearing premise

The load-bearing premise is that the automatically produced sentence-level translations and synthetic target speech give BLEU rewards that are a faithful proxy for good simultaneous interpretation; if those references are biased or too easy, the reinforcement loop optimizes the wrong objective and the quality/latency claim collapses.

What would settle it

Take the released long-form benchmark and replace the automatically generated reference translations with human professional translations of the same source speech; if the model's BLEU/COMET advantage over the supervised baselines shrinks or reverses, the reported gains are tied to the synthetic reference style rather than to interpretation quality. A second check: on partial translations, ask human interpreters to rate whether the model's incremental output is acceptable at each word boundary and compare those ratings with the BLEU process rewards; low correlation would mean the optimization

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

If this is right

  • Training a simultaneous translator reduces to sentence-aligned audio pairs; no word-level aligner or language-specific segmentation heuristic is needed.
  • The same base model can be tuned to different quality/latency trade-offs by adjusting the reward blend, giving deployment flexibility without retraining from scratch.
  • The inner text stream provides a free contemporaneous transcript, so the system doubles as streaming speech-to-text translation.
  • A new input language can be added with less than 1,000 hours of speech, suggesting a path for lower-resource languages.
  • The released 45-hour multilingual benchmark gives the field a common long-form speech-translation evaluation set.

Where Pith is reading between the lines

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

  • [Editorial inference] The tested languages are all relatively monotonic with English; the method's hardest case -- verb-final languages or free-word-order pairs -- remains open. Running the same pipeline on Japanese or Korean to English would directly test whether sentence-level alignments plus process rewards generalize beyond structurally similar pairs.
  • [Editorial inference] Because the BLEU reference is generated by the same automatic translation model used to build training data, part of the reported gain may be style matching rather than semantic fidelity; evaluating against independent human translations would separate the two. The paper itself notes that validation BLEU is far above evaluation BLEU, consistent with this concern.
  • [Editorial inference] The voice-transfer results may depend on the TTS used to synthesize targets with 10-second speaker conditioning; a comparison against a model trained on natural human interpretation speech would show whether the speaker-similarity gains survive outside the synthetic pipeline.
  • [Editorial inference] The authors state they cannot control accent intensity, which is a stated limitation; adding accent-annotated conditioning data would be the direct extension and would make the voice-transfer claim more actionable.

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

5 major / 4 minor

Summary. Hibiki-Zero is a simultaneous speech-to-speech translation system that avoids word-level alignment supervision. It first trains a multistream RQ-Transformer on sentence-level aligned, synthetically built parallel speech data (Whisper transcripts, MADLAD-3B translations, TTS target speech with inserted silences), then uses GRPO with BLEU-based process rewards to reduce latency while preserving translation quality. The paper claims state-of-the-art translation quality, latency, speaker similarity, and naturalness across French, Spanish, Portuguese, and German, as well as fast adaptation to Italian with less than 1000 hours of data. The authors release model weights, inference code, and a new long-form benchmark (Audio-NTREX-4L).

Significance. If the claims hold, the contribution is significant: removing the need for language-specific word-level alignment heuristics would substantially simplify the training pipeline for simultaneous S2ST and could facilitate scaling to more languages. The paper also provides a new benchmark and open-sources the model, which supports reproducibility. However, the central SOTA claim rests on a self-built synthetic benchmark whose construction pipeline closely matches the training data pipeline, and on a BLEU reward computed against the same MT model used to generate the training references. These issues mean the claimed margins may be inflated by distributional match rather than generic improvement. The method itself is novel and plausible, but the evaluation needs independent or at least more comprehensive validation.

major comments (5)
  1. [§4.3, §4.2.3, §4.7] The long-form benchmark Audio-NTREX-4L is constructed with the same synthetic pipeline as the training data: source and target speech are synthesized with TTS, transcripts are produced by Whisper, and translations are produced by MADLAD-3B. The paper admits in §4.7 that the train/valid sets follow the same data-generation process and therefore 'the same translation style as MADLAD-3B.' The largest margins in Table 1 (e.g., French long-form ASR-BLEU 28.7 vs 23.9, speaker similarity 61.3 vs 44.4) may thus reflect distributional matching rather than general simultaneous-interpretation ability. Independent evaluation on natural human speech is needed to support the SOTA claim.
  2. [§2.1, Table 1] The related work cites Seed LiveInterpret 2.0 (Cheng et al.) and Google Real-Time S2ST (Misiunas & Ablavatski) as strong recent systems, but neither is included in Table 1. The abstract claims state-of-the-art performance, yet the comparison set is limited to Seamless and, for French only, Hibiki. Without these baselines, the SOTA claim is not established. The authors should either evaluate these systems on the same benchmark or explicitly justify their exclusion.
  3. [§3.3.1, Eq. (3)] The process reward is BLEU against a sentence-level reference y that, in the training data construction (§4.2.3), is generated by MADLAD-3B. Since the model is trained to replicate MADLAD-3B style, RL is effectively optimizing BLEU to a reference that shares systematic biases with the training target. The paper acknowledges in §4.7 that validation BLEU is inflated by this style match. This circularity means the reported quality/latency trade-offs may not transfer to evaluation with independent human references or alternative MT systems.
  4. [Table 2] The human evaluation does not directly compare against the previous SOTA system (Hibiki) for Spanish, Portuguese, or German; only French has a Hibiki comparison. For the other languages, Hibiki-Zero is compared only to Seamless, whose speaker similarity and naturalness are dramatically lower. Consequently, the claim of superior naturalness and voice transfer over the actual prior SOTA is not demonstrated for most languages. The 50-sample, 20-rater setup also lacks significance tests or confidence intervals for the reported MOS differences.
  5. [§4.3] The long-form evaluation uses TTS-synthesized source speech, not natural human speech. Real simultaneous interpretation involves disfluencies, background noise, accents, and diverse speaking styles; the current benchmark may overestimate real-world performance. The short-form Europarl-ST data is natural speech, but the ASR-BLEU margin over Seamless there is much smaller (e.g., French 1.8 points), which tempers the SOTA claim. The paper would benefit from an additional natural-speech long-form evaluation or at least a discussion of how synthetic evaluation may bias the results.
minor comments (4)
  1. [§3.3.1] The definition of S(t) is slightly off-by-one: 'S(t)=i for ti ≤ t ≤ ti+1' and 'S(t)=n−1 for t > tn' should be clarified, e.g., whether the last sentence index is n−1 or n, since the notation is used in Eq. (3).
  2. [§4.2.2] The data mixture description ('12% in each input language, 50% of English and less than 2% of Italian') does not sum to 100% for the listed languages; specify how the remaining percentage is allocated.
  3. [References] The paper cites MADLAD-3B but the reference is to Kudugunta et al. (2023), which describes MADLAD-400. Please clarify whether a specific model checkpoint is used and cite accordingly.
  4. [Table 1] The table header lists 'ASR BLEU (↑)', 'COMET (↑)', etc., but does not specify which BLEU is computed from the text output vs ASR transcript; although §4.4 explains this, adding a footnote to the table would improve readability.

Circularity Check

0 steps flagged

No significant circularity: central RL claim is evaluated on independent benchmarks; disclosed valid-set overlap is a limitation, not a circular step.

full rationale

The paper's central claim is that RL with BLEU-only process rewards, starting from a sentence-level-aligned base model, yields state-of-the-art simultaneous S2ST quality/latency trade-offs. This claim is supported by evaluations on Europarl-ST (natural human speech with human translations) and Audio-NTREX-4L (human NTREX translations synthesized with external TTS providers). Neither evaluation reuses the training references produced by Whisper/MADLAD-3B/TTS, so the reported ASR-BLEU/COMET gains are not forced by construction. The training reward is BLEU against MADLAD-3B translations, and the evaluation is also BLEU-family, but this is standard objective alignment rather than circularity: the evaluation references are independent of the reward references, and the paper also reports COMET and human MOS. The paper explicitly discloses (Sec. 4.7) that its train and valid sets follow the same data-generation pipeline, which explains high validation BLEU (~60) vs evaluation (~30); this is a model-selection caveat, not a self-definitional derivation step, and it does not taint the independent test sets. Self-citations to Moshi (Defossez et al., 2024) and Hibiki (Labiausse et al., 2025) provide architecture and data-pipeline prior art that is published, open-source, and not used as an unverified uniqueness theorem or ansatz. Missing baselines (Seed LiveInterpret 2.0, Google Real-Time S2ST) and the synthetic nature of long-form source speech are legitimate external-validity concerns, but they are correctness/completeness issues, not circularity. Overall, the derivation chain is self-contained with respect to the stated predictions; no step reduces by definition to its inputs.

Axiom & Free-Parameter Ledger

10 free parameters · 8 axioms · 2 invented entities

The central claim (state-of-the-art simultaneous S2ST without word-level alignments) rests on several hand-chosen hyperparameters (δ, µ, α, nw, c_q, G, T) and several strong assumptions about the synthetic data pipeline: sentence-level correspondence, ASR/MT quality, TTS fidelity, and the absence of reward hacking in GRPO without KL. These do not invalidate the work, but they determine whether the reported results transfer beyond the specific data generation setup.

free parameters (10)
  • δ (silence insertion delay parameter) = 0.5
    Controls how much each target sentence is delayed relative to source in coarse alignment. Chosen by hand, not learned. Affects the base model's ability to explore low-latency policies during RL.
  • µ (pause insertion scale) = 2
    Controls duration of inserted silences at punctuation for coarse alignment. Hand-chosen; affects how natural the synthetic target speech sounds.
  • α (reward balancing term) = 0.4 (main), 0.5 (ablations)
    Balances process BLEU vs outcome BLEU in reward. Tuned by ablation (Fig 4); directly controls the quality/latency trade-off.
  • nw (process-reward word interval) = 8
    Frequency of intermediate reward computation. Ablated (Fig 6); chosen for performance.
  • ϵ (clipping range in PPO objective) = 0.2
    Standard PPO clip; hand-chosen.
  • c0 = 100, cq = 1 for q≥1 = 100:1
    Loss weights balancing text stream vs audio codebook streams; hand-chosen.
  • G (group size for GRPO) = 4
    Number of sampled translations per input; hand-chosen. Affects variance of advantages.
  • T (sequence length in frames) = 1500
    Fixed generation length during RL; hand-chosen.
  • temperature and top-k for generation during RL = 0.8 / 250
    Sampling hyperparameters; hand-chosen.
  • γ in LAAL metric (per-sample scaling) = Δ_source / max(n_gen, n_ref)
    The LAAL formula uses this fitted per-sample scaling constant; not learned but affects reported latency numbers.
axioms (8)
  • domain assumption Sentence-level alignment: X and Y contain the same number of sentences and sentence i in Y is a translation of sentence i in X.
    The entire coarse alignment method rests on this assumption (Sec 3.2.1). If sentence segmentation or correspondence is wrong, the 'sentence-level alignment' is corrupted and RL rewards are computed against wrong prefixes.
  • domain assumption Whisper large-v3 transcripts and Spacy sentence splits are accurate enough for sentence alignment.
    Used to build training data (Sec 4.2.3). Errors in ASR or sentence splitting could produce misaligned sentence pairs.
  • domain assumption MADLAD-3B translations are of sufficient quality to serve as ground-truth references for BLEU rewards.
    The RL reward is BLEU against these translations. If MADLAD-3B has systematic errors or stylistic biases, the model optimizes for those artifacts.
  • domain assumption The TTS model used to synthesize target speech (with controlled pause timing and voice transfer) preserves the linguistic content and allows natural insertion of silences.
    The synthetic target speech Y+ is used both as training target and as evaluation reference audio (through ASR-BLEU/COMET). TTS errors become training noise and eval bias.
  • domain assumption Evaluating translation quality by ASR of generated speech (Whisper medium) introduces no systematic bias favoring the model.
    ASR-BLEU and ASR-COMET rely on Whisper transcriptions of generated audio. If Whisper makes systematic errors on synthetic voices or accented speech, reported BLEU/COMET may be skewed.
  • standard math Mimi codec tokens are sufficient to capture semantic and acoustic content for ST.
    The whole architecture encodes speech into discrete Mimi tokens; if Q=16 RVQ levels lose information needed for high-quality translation or voice transfer, results would degrade. This is standard for audio language models, but untested for extreme multilingual data.
  • domain assumption The multistream RQ-Transformer can learn to translate from non-causal loosely aligned data.
    The paper claims that training on sentence-level coarse alignments with random delays is enough for the base model to learn ST (Sec 3.2). This is empirically validated in the paper, but it is a fragile assumption: if the model ignores the timing signal, RL may not be able to reduce latency without hurting quality (cf. Expt B in Fig 7).
  • domain assumption GRPO without KL regularization does not lead to reward hacking or distribution collapse.
    The paper omits KL in the RL objective. They claim no reward hacking was observed, but no evidence or monitoring is shown. The higher α=0.5 experiments are used as proof of quality/latency control, but without KL, the model could overfit to BLEU-n-gram patterns.
invented entities (2)
  • Hibiki-Zero (system) no independent evidence
    purpose: New model combining multistream architecture with RL-based latency optimization.
    Not a new natural entity; a system. It has falsifiable predictions via benchmark results, but no independent evidence yet beyond this paper.
  • Audio-NTREX-4L benchmark independent evidence
    purpose: New long-form evaluation set built from NTREX text translations and TTS voice synthesis.
    The benchmark itself is a released artifact (claimed) that could be used by others to evaluate independently, but its construction is partly synthetic (TTS-generated) and shares pipeline with training data.

pith-pipeline@v1.3.0-alltime-deepseek · 18270 in / 10329 out tokens · 73658 ms · 2026-08-03T00:52:06.183242+00:00 · methodology

0 comments
read the original abstract

Simultaneous speech translation requires translating source speech into a target language in real-time while handling non-monotonic word dependencies. Traditional approaches rely on supervised training with word-level aligned data, which is difficult to collect at scale and thus depends on synthetic alignments using language-specific heuristics that are suboptimal. We propose Hibiki-Zero, which eliminates the need for word-level alignments entirely. This fundamentally simplifies the training pipeline and enables seamless scaling to diverse languages with varying grammatical structures, removing the bottleneck of designing language-specific alignment heuristics. We first train on sentence-level aligned data to learn speech translation at high latency, then apply a novel reinforcement learning strategy using GRPO to optimize latency while preserving translation quality. Hibiki-Zero achieves state-of-the-art performance in translation accuracy, latency, voice transfer, and naturalness across five X-to-English tasks. Moreover, we demonstrate that our model can be adapted to support a new input language with less than 1000h of speech. We provide examples, model weights, inference code and we release a benchmark containing 45h of multilingual data for speech translation evaluation.

Figures

Figures reproduced from arXiv: 2602.11072 by Alexandre D\'efossez, Neil Zeghidour, Romain Fabre, Tom Labiausse, Yannick Est\`eve.

Figure 1
Figure 1. Figure 1: Architecture of the RQ-Transformer. Figure adapted from Defossez et al. ´ (2024). 0 ... 0 ... ... ... ... ... ... ... ... ... Acoustic tokens Semantic tokens Text tokens Hibiki-Zero stream (speaks) ... ... ... Hibiki-Zero est tellement rapide Hibiki-Zero is so fast 👤 🤖 0 ... 0 ... ... ... ... ... ... ... ... ... ... ... ... Source stream (listens) Acoustic tokens Semantic tokens PAD EPAD Hibiki Zero PAS EP… view at source ↗
Figure 2
Figure 2. Figure 2: Joint sequence modeling. From the source stream, Hibiki-Zero predicts its Inner Monologue text stream, semantic and acoustic tokens. Figure adapted from Labiausse et al. (2025). Defossez et al. ´ (2024), the output of the first quantization level is trained to replicate semantic information obtained from a WavLM self-supervised audio model (Chen et al., 2022). We refer to At,1 as semantic tokens, and At,q≥… view at source ↗
Figure 3
Figure 3. Figure 3: Process rewards method based on BLEU score. We introduce intermediate BLEU score computed on the text output of the model before a given frame t and using the ground-truth translation of the corresponding input sentences processed so far. We combine it with the total output BLEU score using α ∈ [0, 1]. 4. Experiments 4.1. Architectural hyper-parameters The backbone Temporal Transformer of Hibiki-Zero has a… view at source ↗
Figure 4
Figure 4. Figure 4: Influence of hyperparameter α during RL. We plot the BLEU score and text LAAL over training for various α (see Eq. (3)), starting from the same supervised model using nw = 8. New language adaptation. Following our method from Section 4.2.3, we build a small coarse-aligned Italian-to￾English ST dataset containing less than 1000 hours in each language. Starting from the base translation model obtained after … view at source ↗
Figure 7
Figure 7. Figure 7: Alternative configurations. We use α = 0.5 and nw = 8 for all experiments. Experiment (A) uses the full translation of the input speech as reference to compute process rewards instead of sentence-level prefixes as in Equation 3. Experiment (B) performs RL on a supervised model trained with full sentence-delay i.e. δi = di for each input sentence of index i. Experiment (C) performs RL on a supervised model … view at source ↗
Figure 6
Figure 6. Figure 6: Influence of hyperparameter nw during RL. We plot the BLEU score and text LAAL over training for various nw (see Sec. 3.3.2) starting from the same supervised model using α = 0.5. Ablation: Alternative configurations. In [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. NaturalFlow: Reducing Disruptive Pauses for Natural Speech Flow in Simultaneous Speech-to-Speech Translation

    cs.CL 2026-06 unverdicted novelty 6.0

    A fluency-aware optimization framework is introduced to minimize inter-chunk silences in simultaneous speech-to-speech translation by leveraging model-internal signals including linguistic diversity and temporal variability.

  2. Regularized Entropy Information Adaptation with Temporal-Awareness Networks for Simultaneous Speech Translation

    cs.LG 2026-04 unverdicted novelty 5.0

    REINA-SAN and REINA-TAN add temporal context to information-based read/write policies, improving the quality-latency tradeoff in simultaneous speech translation by up to 7.1% on Normalized Streaming Efficiency.

Reference graph

Works this paper leans on

6 extracted references · 5 linked inside Pith · cited by 2 Pith papers

  1. [4]

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O

    doi: 10.48550/ARXIV .2306.12925. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms. 10 Simultaneous Speech-to-Speech Translation Without Aligned Data CoRR, abs/1707.06347, 2017. URL http://arxiv. org/abs/1707.06347. Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Zhang, M., Li, Y . K., Wu, Y ., and ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.12925 2017

  2. [5]

    findings-acl.828/

    URL https://aclanthology.org/2025. findings-acl.828/. Yang, D., Tian, J., Tan, X., Huang, R., Liu, S., Chang, X., Shi, J., Zhao, S., Bian, J., Wu, X., et al. Uniaudio: An au- dio foundation model toward universal audio generation. arXiv preprint arXiv:2310.00704, 2023. Yu, D., Zhao, Y ., Zhu, J., Xu, Y ., Zhou, Y ., and Zong, C. Simulpl: Aligning human pr...

    Pith/arXiv arXiv 2025

  3. [6]

    URL https: //doi.org/10.48550/arXiv.2509.08753

    doi: 10.48550/ARXIV .2509.08753. URL https: //doi.org/10.48550/arXiv.2509.08753. Zhang, S., Fang, Q., Guo, S., Ma, Z., Zhang, M., and Feng, Y . Streamspeech: Simultaneous speech-to-speech translation with multi-task learning. In Ku, L., Mar- tins, A., and Srikumar, V . (eds.),Proceedings of the 62nd Annual Meeting of the Association for Computa- tional Li...

    doi:10.48550/arxiv.2509.08753 2024

  4. [213]

    Rubenstein, P

    URL https://doi.org/10.18653/v1/ 2020.emnlp-main.213. Rubenstein, P. K., Asawaroengchai, C., Nguyen, D. D., Bapna, A., Borsos, Z., de Chaumont Quitry, F., Chen, P., Badawy, D. E., Han, W., Kharitonov, E., Muckenhirn, H., Padfield, D., Qin, J., Rozenberg, D., Sainath, T. N., Schalkwyk, J., Sharifi, M., Ramanovich, M. T., Tagliasac- chi, M., Tudor, A., Veli...

    Pith/arXiv arXiv 2020

  5. [2023]

    doi: 10.48550/ARXIV .2312.05187. Chen, S., Wang, C., Chen, Z., Wu, Y ., Liu, S., Chen, Z., Li, J., Kanda, N., Yoshioka, T., Xiao, X., Wu, J., Zhou, L., Ren, S., Qian, Y ., Qian, Y ., Wu, J., Zeng, M., Yu, X., and Wei, F. Wavlm: Large-scale self-supervised pre-training for full stack speech processing.IEEE J. Sel. Top. Signal Process., 2022. Cheng, S., Bao...

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.05187 2022

  6. [2025]

    URL https: //doi.org/10.48550/arXiv.2506.10910

    doi: 10.48550/ARXIV .2506.10910. URL https: //doi.org/10.48550/arXiv.2506.10910. Rei, R., Stewart, C., Farinha, A. C., and Lavie, A. COMET: A neural framework for MT evaluation. InProceed- ings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, pp. 2685– 2702, 2020. doi: 10.18653/V1/2020.EMNLP-MAIN

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.10910 2020