Pith. sign in

REVIEW 2 major objections 6 minor 37 references

With verifiable rewards, reinforcement learning adapts audio-language models to code-switched ASR on a tenth of the data that full supervised fine-tuning needs, and the gains hold on real human recordings.

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-4.5

2026-07-12 07:14 UTC pith:W2V5LJRU

load-bearing objection Solid empirical recipe for CS-ASR with clean ablations; the 10%-beats-100% headline is real under their setup but confounded by full-decoder GRPO vs LoRA capacity. the 2 major comments →

arxiv 2607.02757 v1 pith:W2V5LJRU submitted 2026-07-02 cs.CL cs.SD

Reinforcement Learning for Data-Efficient Code-Switched ASR

Ziwei Ye , Peter Vickers This is my paper
classification cs.CL cs.SD
keywords reinforcement learningcode-switchingautomatic speech recognitionreward designgroup relative policy optimizationscript fidelityspeech-LLM adaptationdata efficiency
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.

Code-switched speech is common, but large audio-language models still break at language boundaries: they translate one language into another, drop mixed spans, or emit characters from the wrong writing system. This paper treats transcription quality as a verifiable reward problem and adapts a public speech-LLM with group-relative policy optimization, using character error rate plus a script-fidelity bonus that penalizes disallowed writing systems, plus an optional two-pass draft-and-refine step during training. Trained only on synthetic TTS code-switched audio, the method with 10% of the data matches LoRA supervised fine-tuning on the full set across ten language pairs, and with 20% it outperforms it, with the biggest lifts on typologically distant pairs. The error-rate reward largely kills wholesale translation failures; the script reward separately cuts script contamination without hurting accuracy. Those improvements transfer zero-shot to a human-recorded code-switching corpus.

Core claim

RLVR with GRPO, a CER reward, a script-fidelity reward, and two-pass draft refinement, trained on only 10% of TTS code-switched data, matches full-data LoRA SFT on human-read code-switched test sets across ten language pairs, and at 20% data beats it, with largest gains on distant pairs and zero-shot transfer to human-recorded SwitchLingua.

What carries the argument

Reinforcement learning with verifiable rewards (RLVR) via group relative policy optimization (GRPO): sample several candidate transcripts, score them with −CER plus a small binary script-fidelity bonus for allowed writing systems, normalize advantages within the group, and update; optionally run a second GRPO pass that conditions on the best draft to train self-correction.

Load-bearing premise

The headline data-efficiency claim rests on comparing full-decoder RL updates on a small TTS subset against low-rank LoRA supervised fine-tuning on the full set, so capacity and objective are not equalized.

What would settle it

Equalize trainable capacity (full decoder vs full decoder, or LoRA vs LoRA) and re-run the 10% RLVR versus 100% SFT comparison on CS-FLEURS read_test and SwitchLingua; if RLVR no longer matches or beats SFT, or if the script reward no longer cuts wrong-script rate without raising CER, the central efficiency claim fails.

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

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

2 major / 6 minor

Summary. The paper proposes a practical RLVR recipe for adapting audio-language models to code-switched ASR: GRPO with a CER reward, a binary script-fidelity (SHR) reward that penalizes disallowed Unicode scripts, and a training-time two-pass draft-and-refinement procedure. Using Qwen2-Audio-7B-Instruct as a fixed testbed, models are trained only on TTS CS-FLEURS XTTS-TRAIN subsets (10%/20%/100%) across 10 language pairs and evaluated on human-read CS-FLEURS and zero-shot on human-recorded SwitchLingua. The central empirical claim is that RLVR with 10% of the TTS data matches LoRA SFT trained on 100% of the same data (and at 20% outperforms it), with largest gains on typologically distant pairs; ablations attribute translation-error collapse to the CER reward and residual script-contamination reduction to the SHR reward, with gains transferring across acoustic domains.

Significance. If the comparison is interpreted carefully, this is a useful, reproducible recipe paper for data-efficient CS-ASR adaptation of speech-LLMs. Strengths include multi-seed reporting (n=3) with means and stds in Table 1, a clean ablation chain isolating CER-only vs refinement vs SHR (Figure 2), explicit verifiable rewards rather than opaque preference models, and zero-shot transfer from TTS training to human-recorded SwitchLingua. The script-fidelity reward is a simple, falsifiable consistency signal well matched to observed failure modes (translation and wrong-script contamination). The work is not claiming SOTA ASR; it is a controlled study of reward design and data efficiency on a public model, which is appropriate and valuable for the community.

major comments (2)
  1. The headline data-efficiency claim (abstract, §5, Table 1) compares full-decoder GRPO updates on 10–20% of XTTS-TRAIN against LoRA SFT (r=64) on 100% of the same data. Section 4.3 states both freeze the audio encoder, but RLVR “updates all decoder parameters” while LoRA updates only low-rank adapters, and gradient-step budgets also differ substantially (132–264 vs 5,176). Without a full-parameter supervised baseline and/or a LoRA-parameterized GRPO run, the result confounds sequence-level RLVR with unrestricted decoder capacity and update budget. The central claim remains defensible as “RLVR on a small TTS subset can match LoRA-on-full,” but the stronger framing “data-efficient adaptation of the RLVR recipe itself” is not isolated. Please add at least one matched-capacity control, or reframe the claim and abstract to state the comparison exactly as run.
  2. Table 1 and §6.1 show that two-pass refinement is not uniformly helpful: it improves several pairs on CS-FLEURS but consistently degrades ara-eng on both CS-FLEURS and SwitchLingua, and on SwitchLingua refinement alone does not beat single-pass CER at 10% data. The abstract and contribution bullets present draft-and-refinement as a core part of the recipe. Please qualify when refinement helps vs hurts (especially script-boundary-ambiguous pairs), and either condition its use or present it as optional rather than a default component of the method.
minor comments (6)
  1. Naming: “SHR” is used both for Script Hallucination Rate (evaluation metric, §4.4) and for the script-fidelity reward (Section 3.3). Consider “script-fidelity reward / SF” vs “SHR (rate)” to avoid conflating reward and metric.
  2. Figure 1 is effective but dense; the intermediate “Format-only baseline” and “LoRA SFT” lines are hard to parse as currently laid out. A cleaner three-column layout (reference / failure / success) would help.
  3. Section 3.3: state explicitly how allowed scripts S_L1 ∪ S_L2 are obtained (Unicode script property tables? language-specific exceptions for shared characters?) so the reward is fully reproducible for new pairs.
  4. Table 1 header “CER+SHR+refine” etc. is clear, but the parenthetical SHR values could be labeled in the caption (e.g., “CER (SHR)”) so readers do not need to infer the second number.
  5. Conclusion claims “∼40× fewer gradient steps”; Table 1 supports a large step reduction, but the exact factor depends on which RLVR column is compared (132 vs 5,176 ≈ 39×; 264 vs 5,176 ≈ 20×). Align the prose with the specific setting cited.
  6. Minor prose: “These issues can be amplified at code-switch boundaries” (§1); “On SwitchLingua, 37%… are fully translated” (§3.4) — useful statistics; consider moving a short failure-mode table into the main text for scanability.

Circularity Check

0 steps flagged

No circular derivation: empirical RLVR recipe evaluated on held-out CER/SHR; rewards guide policy learning, they do not redefine the reported metrics by construction.

full rationale

This is an empirical methods paper, not a first-principles derivation. The load-bearing claim is that GRPO with a CER reward, an optional script-fidelity (SHR) bonus, and optional two-pass draft refinement, trained on 10–20% of CS-FLEURS XTTS-TRAIN, matches or beats LoRA SFT trained on 100% of the same split on held-out CS-FLEURS read_test and zero-shot SwitchLingua (Table 1, §5). That claim is supported by multi-seed experimental comparison, not by algebraic identity or fitted-parameter renaming. Optimizing −CER (plus a binary Unicode-script bonus) and then reporting CER/SHR is standard RLVR / sequence-level training: advantages are z-scored within sampled groups (Eq. 1–2, §3.2–3.3), the policy must still generate better transcripts on new audio, and evaluation uses single-pass greedy decoding on held-out human-read and human-recorded sets. The SHR reward is an independent script-membership check, not a redefinition of CER. There is no self-definitional loop, no parameter fitted on a subset then re-reported as a prediction of a closely related quantity, no load-bearing uniqueness theorem imported from the present authors, and no ansatz smuggled in via self-citation. Capacity/update-budget confounds (full-decoder GRPO vs LoRA adapters) are experimental-design concerns, not circularity. Score 0 with empty steps is the correct outcome.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central claim is empirical and rests on standard RL/ASR machinery plus a few hand-chosen knobs and modeling choices. No new physical entities. Load-bearing free parameters are the SHR bonus weight and GRPO sampling/optimization settings; axioms are domain assumptions about verifiable rewards, Unicode script sets as allowed writing systems, and the fairness of the LoRA comparison setup.

free parameters (4)
  • β_sf (script fidelity bonus)
    Hand-set to 0.05; scales the binary Script(ŷ,C) term in the composite reward and therefore how strongly clean-script candidates win within each GRPO group.
  • G (group size / completions per prompt)
    Set to 8; controls the relative-advantage estimate quality and compute per step.
  • clip ε, learning rate, temperature
    ε=0.2, lr=1e-6, τ=1.0 chosen as training hyperparameters that affect policy update stability and exploration; not derived from data but required for the reported runs.
  • data subsample fractions (10%, 20%)
    Fixed fractions (n=2310 / 4625) used to support the data-efficiency claim; results depend on these specific subsets across 10 pairs.
axioms (5)
  • domain assumption Group Relative Policy Optimization with z-scored within-group advantages is a valid critic-free RL update for sequence-level ASR rewards.
    Invoked throughout Section 3.2 and Eq. (1); taken from DeepSeekMath/GRPO literature without re-derivation.
  • domain assumption Character error rate against a reference is a sufficient verifiable reward to eliminate translation-mode failures in code-switched ASR.
    Section 3.3 and analysis in 6.1; CER is treated as the primary quality signal.
  • ad hoc to paper Allowed writing systems for a language pair equal the union of Unicode scripts of L1 and L2 (plus punctuation/digits/whitespace).
    Defines Script(ŷ,C) and SHR in Section 3.3; a modeling choice that may mis-handle loanwords, romanization, or mixed orthographies.
  • domain assumption TTS-synthesized CS-FLEURS XTTS speech is an adequate training domain for policies evaluated on human-read CS-FLEURS and human-recorded SwitchLingua.
    Training setup in Section 4.1; transfer is tested but remains an assumption of the data-efficiency recipe.
  • domain assumption Freezing the Whisper-based audio encoder and updating only the LLM decoder is sufficient for CS adaptation under both RLVR and LoRA.
    Section 4.3 training details; encoder adaptation is not explored.
invented entities (2)
  • Script fidelity (SHR) reward no independent evidence
    purpose: Binary bonus that penalizes characters outside the allowed Unicode scripts of the language pair to reduce script contamination at CS boundaries.
    Introduced in Section 3.3 as R = −CER + β_sf · Script(ŷ,C); evaluation-time SHR is its counterpart. Independent evidence is limited to the paper’s own ablations showing SHR drop without CER degradation.
  • Two-pass draft-and-refinement GRPO procedure no independent evidence
    purpose: Train the model to sample drafts, update on them, then refine the best draft conditioned on audio plus draft within one training iteration.
    Algorithm 1; inspired by self-refine literature but specialized as a training-time GRPO loop for CS-ASR. Inference remains single-pass.

pith-pipeline@v1.1.0-grok45 · 14654 in / 3839 out tokens · 42591 ms · 2026-07-12T07:14:03.331164+00:00 · methodology

0 comments
read the original abstract

Audio-language models can be prompted for code-switched speech, but their decoding is not optimized for code-switching and often fails at language boundaries. We propose a practical reinforcement learning with verifiable rewards recipe for data-efficient adaptation of audio-language models to code-switched ASR using group relative policy optimization, combining an error rate reward with a script fidelity reward that penalizes wrong writing systems and a two-pass draft-and-refinement procedure. Using Qwen2-Audio as a reproducible testbed across 10 language pairs, training on only TTS code-switched speech, we show that RLVR with 10% of the data matches LoRA supervised fine-tuning trained on the full dataset, with the largest gains on typologically distant pairs. The error rate reward eliminates translation errors while the script fidelity reward separately reduces script contamination without degradation. These gains transfer zero-shot to a human-recorded code-switching corpus.

Figures

Figures reproduced from arXiv: 2607.02757 by Peter Vickers, Ziwei Ye.

Figure 1
Figure 1. Figure 1: Chinese-English code-switch sample in SwitchLingua • We apply Reinforcement Learning with Verifiable Re￾wards (RLVR) [10] on code-switched ASR, using rewards based on CER and a script fidelity reward to directly en￾courage correct writing systems at code-switch boundaries. • We introduce a training-time two-pass draft and refine￾ment procedure that conditions a second decoding pass on the best draft to enc… view at source ↗
Figure 2
Figure 2. Figure 2: Translation and wrong-script rates across the 10% RLVR ablation chain on non-Latin-script samples. SL = Switch￾Lingua, CF = CS-FLEURS. CER reward eliminates translation. The SHR reward is needed to reduce script contamination. 5. Results CS-FLEURS human-read speech [PITH_FULL_IMAGE:figures/full_fig_p004_2.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

37 extracted references · 4 linked inside Pith

  1. [1]

    listen again and fix

    Introduction Code-switching is a natural communicative behavior for the majority of the world’s multilingual population [1, 2]. De- spite its prevalence, code-switched automatic speech recogni- tion (CS-ASR) remains challenging due to the need to model rapidly shifting acoustic and linguistic cues, compounded by the scarcity of labeled code-switched train...

  2. [2]

    Recent work incorporates explicit language tag sig- nals and language-aware decoding [14]

    Related work Code-switched ASR.Code-switching ASR must jointly model acoustics and rapidly shifting language priors, and is challenged by language confusion errors and limited labeled code-switched data [13]. Recent work incorporates explicit language tag sig- nals and language-aware decoding [14]. FLEURS provides a standardized multilingual benchmark [15...

    Pith/arXiv arXiv 2026

  3. [3]

    The audio contains speech mixing Japanese and English

    Method 3.1. Problem Setup Given an audio segmentx, a language contextC, and a refer- ence transcripty ∗, the model produces a transcriptˆy∼π θ(· | x,C). Our goal is to improve both transcription accuracy and code-switch fidelity. The language contextCconsists of a instruction specifying the language pair (“The audio contains speech mixing Japanese and Eng...

  4. [4]

    Datasets CS-FLEURS.CS-FLEURS [11] is a massively multilingual code-switched speech dataset built on FLEURS [15], covering 52 languages across 113 code-switched pairs

    Experimental setup 4.1. Datasets CS-FLEURS.CS-FLEURS [11] is a massively multilingual code-switched speech dataset built on FLEURS [15], covering 52 languages across 113 code-switched pairs. Code-switched text is generated with an align-then-swap procedure that re- places 30% of content words using word alignments from Awe- someAlign [25], yielding consis...

  5. [5]

    Even at 10% data, RLVR matches or surpasses LoRA trained on the full dataset

    Results CS-FLEURS human-read speech.Table 1 shows that with only 20% of the training data, RLVR with two-pass refinement and the SHR reward achieves a micro-averaged CER of0.147, outperforming 100% LoRA SFT (0.159) despite using5×less data. Even at 10% data, RLVR matches or surpasses LoRA trained on the full dataset. Two-pass refinement provides a clear g...

  6. [6]

    seventy km

    Analysis and Discussion 6.1. Do the rewards address code-switching failures? To isolate the contribution of each reward component, we mea- sure translation rate (non-English portion rendered as English) and wrong-script rate (characters from an unrelated writing sys- tem) across the ablation chain on non-Latin-script samples. Fig- ure 2 summarizes the res...

  7. [7]

    Conclusion We showed that RLVR with GRPO enables highly data-efficient adaptation of speech-LLMs to code-switched ASR. With only 10% of the training data and∼40×fewer gradient steps, RLVR matches LoRA SFT trained on the full dataset across 10 lan- guage pairs, and at 20% outperforms it, with the largest gains on typologically distant pairs where code-swit...

  8. [8]

    Additionally, LLMs were used for spellchecking and suggesting edits to further improve the flu- ency and conciseness of the paper during the drafting process

    Generative AI Use Disclosure LLMs were used to produce a fraction of the source codes for experiment orchestration. Additionally, LLMs were used for spellchecking and suggesting edits to further improve the flu- ency and conciseness of the paper during the drafting process

  9. [9]

    Sometimes I’ll start a sentence in Spanish y termino en español: Toward a typology of code-switching,

    S. Poplack, “Sometimes I’ll start a sentence in Spanish y termino en español: Toward a typology of code-switching,”Linguistics, vol. 18, no. 7–8, pp. 581–618, 1980

    1980

  10. [10]

    A survey of code-switching: Linguistic and social perspectives for language technologies,

    A. S. Do ˘gruöz, S. Sitaram, B. E. Bullock, and A. J. Toribio, “A survey of code-switching: Linguistic and social perspectives for language technologies,” inProceedings of the 59th Annual Meet- ing of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP), 2021, pp. 1654–1666

    2021

  11. [11]

    The decades progress on code-switching research in NLP: A systematic sur- vey on trends and challenges,

    G. Winata, A. F. Aji, Z. X. Yong, and T. Solorio, “The decades progress on code-switching research in NLP: A systematic sur- vey on trends and challenges,” inFindings of the Association for Computational Linguistics (ACL), 2023, pp. 2936–2978

    2023

  12. [12]

    A survey of code-switched speech and language processing,

    S. Sitaram, K. R. Chandu, S. K. Rallabandi, and A. W. Black, “A survey of code-switched speech and language processing,”arXiv preprint arXiv:1904.00784, 2019

    Pith/arXiv arXiv 1904

  13. [13]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” 2022

    2022

  14. [14]

    wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33. Curran Associates, Inc., 2020, pp. 12 449–12 460

    2020

  15. [15]

    Qwen-Audio: Advancing universal audio under- standing via unified large-scale audio-language models,

    Y . Chuet al., “Qwen-Audio: Advancing universal audio under- standing via unified large-scale audio-language models,” 2023

    2023

  16. [16]

    Sequence level training with recurrent neural networks,

    M. Ranzato, S. Chopra, M. Auli, and W. Zaremba, “Sequence level training with recurrent neural networks,” inInternational Conference on Learning Representations (ICLR), 2016

    2016

  17. [17]

    Reducing language confusion for code-switching speech recognition with token-level language diarization,

    H. Liu, H. Xu, L. P. Garcia, A. W. H. Khong, Y . He, and S. Khu- danpur, “Reducing language confusion for code-switching speech recognition with token-level language diarization,” inIEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023

    2023

  18. [18]

    DeepSeekMath: Pushing the limits of mathemati- cal reasoning in open language models,

    Z. Shaoet al., “DeepSeekMath: Pushing the limits of mathemati- cal reasoning in open language models,” 2024

    2024

  19. [19]

    CS-FLEURS: A massively multilingual and code- switched speech dataset,

    B. Yan, I. Hamed, S. Shimizu, V . Lodagala, W. Chen, O. Iakovenko, B. Talafha, A. Hussein, A. Polok, K. Chang, D. Klement, S. Althubaiti, P. Peng, M. Wiesner, T. Solorio, A. Ali, S. Khudanpur, S. Watanabe, C.-C. Chen, Z. Wu, K. Benhar- rak, A. Diwan, S. Cornell, E. Yeo, K. Choi, C. Carvalho, and K. Rosero, “CS-FLEURS: A massively multilingual and code- sw...

    2025

  20. [20]

    SwitchLingua: The first large-scale multilingual and multi-ethnic code-switching dataset,

    P. Xie, X. Liu, Y . Bie, T. W. Chan, Y . Song, Y . Wang, H. Chen, and K. Chen, “SwitchLingua: The first large-scale multilingual and multi-ethnic code-switching dataset,” inAdvances in Neu- ral Information Processing Systems 38 (NeurIPS), Datasets and Benchmarks Track, 2025

    2025

  21. [21]

    Code- switching in automatic speech recognition: The issues and future directions,

    M. B. Mustafa, M. A. Yusoof, H. K. Khalaf, A. A. R. M. Abushariah, M. L. M. Kiah, H. N. Ting, and S. Muthaiyah, “Code- switching in automatic speech recognition: The issues and future directions,”Applied Sciences, vol. 12, no. 19, p. 9541, 2022

    2022

  22. [22]

    Text-derived language identity incorpora- tion for end-to-end code-switching speech recognition,

    Q. Wang and H. Li, “Text-derived language identity incorpora- tion for end-to-end code-switching speech recognition,” inPro- ceedings of the 6th Workshop on Computational Approaches to Linguistic Code-Switching. Singapore: Association for Compu- tational Linguistics, dec 2023, pp. 33–42

    2023

  23. [23]

    FLEURS: Few-shot learning evaluation of universal representations of speech,

    A. Conneau, M. Ma, S. Khanuja, Y . Zhang, V . Axelrod, S. Dalmia, J. Riesa, C. Rivera, and A. Bapna, “FLEURS: Few-shot learning evaluation of universal representations of speech,” 2022

    2022

  24. [24]

    Instruction-following speech recognition,

    C.-I. J. Laiet al., “Instruction-following speech recognition,” 2023

    2023

  25. [25]

    On exposure bias, hallucination and domain shift in neural machine translation,

    C. Wang and R. Sennrich, “On exposure bias, hallucination and domain shift in neural machine translation,” inProceedings of the 58th Annual Meeting of the Association for Computational Lin- guistics (ACL), 2020, pp. 3544–3552

    2020

  26. [26]

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

    R. Prabhavalkaret al., “Minimum word error rate training for attention-based sequence-to-sequence models,” inProceedings of the IEEE International Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), 2018

    2018

  27. [27]

    Optimizing expected word error rate via sampling for speech recognition,

    M. Shannonet al., “Optimizing expected word error rate via sampling for speech recognition,” inProc. INTERSPEECH 2017, 2017

    2017

  28. [28]

    Reinforcement learning with verifiable rewards implicitly incentivizes correct reasoning in base LLMs,

    X. Wenet al., “Reinforcement learning with verifiable rewards implicitly incentivizes correct reasoning in base LLMs,” 2025

    2025

  29. [29]

    DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning,

    D. Guoet al., “DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning,” 2025

    2025

  30. [30]

    Group rel- ative policy optimization for speech recognition,

    P. G. Shivakumar, Y . Gu, A. Gandhe, and I. Bulyko, “Group rel- ative policy optimization for speech recognition,”arXiv preprint arXiv:2509.01939, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Omni-R1: Do you really need audio to fine-tune your audio LLM?

    A. Rouditchenko, S. Bhati, E. Araujo, S. Thomas, H. Kuehne, R. Feris, and J. Glass, “Omni-R1: Do you really need audio to fine-tune your audio LLM?”arXiv preprint arXiv:2505.09439, 2025

    arXiv 2025

  32. [32]

    Self-refine: Iterative refinement with self-feedback,

    A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegr- effe, U. Alon, N. Dziri, S. Prabhumoye, Y . Yang, S. Gupta, B. P. Majumder, K. Hermann, S. Welleck, A. Yazdanbakhsh, and P. Clark, “Self-refine: Iterative refinement with self-feedback,” inAdvances in Neural Information Processing Systems 36 (NeurIPS), 2023

    2023

  33. [33]

    Word alignment by fine-tuning embed- dings on parallel corpora,

    Z.-Y . Dou and G. Neubig, “Word alignment by fine-tuning embed- dings on parallel corpora,” inProceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics (EACL). Association for Computational Linguistics, 2021, pp. 2112–2128

    2021

  34. [34]

    XTTS: a massively multilingual zero-shot text-to-speech model,

    E. Casanova, K. Davis, E. Gölge, G. Göknar, I. Gulea, L. Hart, A. Aljafari, J. Meyer, R. Morais, S. Olayemi, and J. Weber, “XTTS: a massively multilingual zero-shot text-to-speech model,” inInterspeech 2024, 2024, pp. 4978–4982

    2024

  35. [35]

    LoRA: Low-rank adaptation of large lan- guage models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large lan- guage models,” inInternational Conference on Learning Repre- sentations (ICLR), 2022

    2022

  36. [36]

    Qwen2 technical report,

    A. Yang, B. Yang, B. Hui, B. Zheng, B. Yu, C. Zhou, C. Li, C. Li, D. Liu, F. Huanget al., “Qwen2 technical report,”arXiv preprint arXiv:2407.10671, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    DeepSpeed: System optimizations enable training deep learning models with over 100 billion parameters,

    J. Rasley, S. Rajbhandari, O. Ruwase, and Y . He, “DeepSpeed: System optimizations enable training deep learning models with over 100 billion parameters,” inProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ACM, 2020, pp. 3505–3506

    2020