Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Faster Machine Translation Ensembling with Reinforcement Learning and Competitive Correction

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that candidate selection, not fusion, is the bottleneck in neural machine translation ensembling, and that a Deep Q-Network trained on fusion-block rewards — plus a competitive correction step — produces better…

desk verdict SmartGen has a promising RL-based approach to MT ensembling, but the group-selection claim rests on an untested independence assumption and the evaluation has clear gaps. read the letter →

arxiv 2501.15219 v1 pith:RM2MY7XE submitted 2025-01-25 cs.CL

Kritarth Prasad , Mohammadi Zaki , Pratik Singh , Pankaj Wasnik This is my paper · ORCID
classification cs.CL
keywords machinetranslationensemblingdeepQ-networkcandidateselectionreinforcementlearningfusionblockcompetitivecorrectionEnglish-Hindineural
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

This paper claims the bottleneck in neural machine translation ensembling is candidate selection, not the fusion model itself, and that casting selection as a reinforcement learning problem fixes it. SmartGen uses a Deep Q-Network to pick, for each source sentence, a small fixed set of K candidate translations from a pool of L models, passing only those to an encoder-decoder fusion block, and trains the picker with the fusion block's BLEU score as reward so selection and fusion are optimized jointly. The follow-on SmartGen++ adds a Competitive Correction Block that scores the chosen candidates with a reward model and, when a weak one is detected, rewrites it with a large language model conditioned on the rejected candidates. On English-Hindi and Hindi-English benchmarks, the paper reports that SmartGen is about 2.31 times faster than ranker-based ensembling and that SmartGen++ beats the best ranker baseline by 4.48% BLEU. If the claims hold, ensembling can be made both cheaper and better by aligning the selection objective with the fusion block's output.

What carries the argument

The load-bearing mechanism has three parts. First, a Deep Q-Network with a ResNet backbone maps each source sentence (state) to Q-values over the L candidate model indices (actions); the group passed to the fusion block is the top-K Q-values, and the training reward is the sacreBLEU of the fusion block's output, stored in an experience replay buffer. Second, a reward model rθ(x, y), trained with a modified preference loss over sets of four preferred and all rejected candidates, scores every candidate and reveals weak selected candidates by the margin between their rewards. Third, a correction block takes any candidate whose reward margin falls below a threshold τ, plus the rejected candidates and their scores, and asks an enhancer LLM G to produce a replacement translation; that replacement enters the fusion set. The paper's guiding identity is that a candidate group is only as good as what the fusion block does with it, so the selector's objective is the fusion block's output quality, not an independently learned ranking.

What would settle it

On a held-out test set, compute the brute-force optimal K-subset per sentence by fusing all C(L,K) combinations and compare it with the DQN's top-K choice: if the DQN rarely matches the oracle yet trails it by a large BLEU margin, the additive group assumption is the cause. A direct test is to train a DQN whose actions are entire K-subsets; if that variant does not beat top-K selection, group non-separability is not the binding constraint.

Watch

Extended reading notes

Core claim

The paper's central claim is that the usual 'score candidates, keep the top K, fuse' recipe is doubly flawed: the scorer is trained independently of the fusion block, and the fused output is dragged down by the worst candidate in the selected group. SmartGen redefines the task as an MDP in which the state is the source sentence, the actions are candidate model indices, and the reward is the sacreBLEU of the translation produced by the fusion block from the selected group; the DQN selects the group by taking the top-K Q-values. SmartGen++ adds a Competitive Correction Block whose reward model, trained on sets of preferred and rejected candidates, flags weak selected candidates by their reward margin, and an LLM then rewrites them using the rejected candidates as context. The experiments are meant to show that joint training improves quality over rankers that ignore the fusion block, that the correction step recovers quality lost to weak candidates, and that selecting only K models cuts the ensemble's inference time from Ω(L) to O(K).

Load-bearing premise

The DQN selects the group by taking the K candidates with the highest individual Q-values, which assumes the value of a group equals the sum of its parts; if the fusion block's gain depends on combinatorial interactions between candidates, this top-K rule can miss the optimal group even when the Q-values are perfectly learned.

Editorial extensions

If this is right

  • Ensembling cost drops from running all L candidate models to running only the K selected ones per sentence — an Ω(L) to O(K) reduction in model inference time.
  • Because the DQN's reward is the fusion output's sacreBLEU, the selector is trained to pick groups that fuse well rather than individually strong translations, closing the gap identified in the motivation.
  • The Competitive Correction Block decouples candidate improvement from selection: any select-and-fuse ensemble can bolt it on to repair weak candidates, which the paper notes is of independent interest.
  • Joint training and correction together produce the reported gains: SmartGen++ tops the BLEU, chrF++, and Comet tables on most English-Hindi and Hindi-English test sets, while the DQN is the fastest ranker in the comparison.

Reading between the lines

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

  • If the paper's central claims hold, the top-K Q-value rule is the likely weak point, because the paper's own brute-force experiment shows the optimal triplet changes per sentence, implying group value is not separable into per-candidate values; a set-valued action space or beam search over K-subsets should close part of the gap to the oracle.
  • The reward model, trained on preferred and rejected candidate sets, could serve as a reference-free quality judge; the ablation compares reward scores to GPT-based scoring, but the paper stops short of proposing the reward model as a metric itself.
  • Because the Competitive Correction Block is described as decoupled from the DQN, a natural extension is to attach it to PairRanker or SimCLS; if it lifts those baselines as much as it lifts SmartGen, the correction step is doing the heavy lifting rather than the RL selector.
  • The paper admits K is fixed; extending the DQN to emit a per-sentence group size would exploit the observed sentence-to-sentence variation in optimal triplet composition, though the paper does not pursue variable K.
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 / 6 minor

Summary. The paper proposes SmartGen and SmartGen++, two methods for ensembling machine translation systems. SmartGen uses a Deep Q-Network (DQN) to select a small, fixed number of candidate translations from a pool of L models, which are then passed to a fusion block; the DQN is trained using the BLEU score of the final fused output as reward. SmartGen++ adds a Competitive Correction Block (CCB) that uses a learned reward model and an LLM to rewrite weak selected candidates before fusion. The authors evaluate on English-Hindi and Hindi-to-English translation across several benchmarks, reporting BLEU, chrF++, and COMET scores, and claim reduced inference cost compared to ranker-based baselines while achieving competitive or better quality.

Significance. If the claims are substantiated, the paper addresses a relevant practical problem: the high inference cost of ensembling multiple NMT models. The idea of training the selection module with feedback from the fusion block is sensible, and the correction block is an interesting addition. The paper also provides a useful comparison of several ranking methods for NMT ensembling. However, the central contributions are weakened by a mismatch between the stated goal of selecting optimal groups and the actual DQN action space, by a factual contradiction in the reported correction results, and by the lack of statistical validation for very small BLEU gains. These issues must be resolved before the claims of state-of-the-art performance can be accepted.

major comments (4)
  1. [Section 3.1 and Figure 2] The DQN selects the candidate set bT by taking the top-K per-action Q-values, but the reward is the BLEU of the fusion output over the entire selected set, so the per-candidate Q-value is not a well-defined standalone quality score because the reward for selecting candidate a depends on which other candidates are simultaneously selected. The paper's motivation (Figure 2) shows that the brute-force optimal triplet varies across sentences and no fixed subset dominates, which is consistent with combinatorial interactions among candidates, yet no experiment compares SmartGen's selected groups against the brute-force oracle used in that figure. The claim of "identifying optimal groups to pass to the fusion block" therefore rests on an untested additivity assumption; if the DQN is in effect only ranking individual candidates, the group-optimization contribution is not demonstrated.
  2. [Section 4.6 and Table 4] The sentence "Our findings indicate that the CCB block generates alternative translations that surpass the current translation on BLEU score metrics" is contradicted by Table 4 for the Hindi-to-English direction: SmartGen's average BLEU is reported as 42.73, while all four LLMs used as the enhancer G in the CCB produce 42.28–42.30, which is below the current translation. This undermines the claim that CCB improves candidate quality in both directions and needs to be corrected, either by reporting separate results per dataset or by revising the claim.
  3. [Section 4.6 and Table 5] Using the reward model as the judge in Table 5 is partially circular. The reward model is trained (Section 3.2, Eq. 1) with preferred responses that include the top-3 BLEU candidates, and SmartGen is trained with BLEU-based rewards; therefore the higher reward for SmartGen/SmartGen++ over LLM-Blender may reflect the RM's inherited BLEU preferences rather than genuine quality improvements. The paper should present reference-based metrics as the primary evidence and treat the reward comparison as an auxiliary, or show agreement with an independent judge such as GPT-4 scores.
  4. [Tables 2 and 3] Tables 2 and 3 report a single run without error bars or significance tests, and many of the reported gains are tiny (e.g., SmartGen improves over LLM-Blender by 0.22 BLEU on average for English-to-Hindi and by 0.21 for Hindi-to-English). Given the noise typical of BLEU on test sets of a few thousand sentences, such small differences cannot be distinguished from chance. The paper should report variance across multiple runs or bootstrap confidence intervals, and avoid claims of "SOTA performance" on the basis of marginal average improvements.
minor comments (6)
  1. [Section 3.1] The notation "arg, T opKa, Q(s, a)" in Section 3.1 appears malformed; it should presumably denote selecting the K actions with the highest Q-values.
  2. [Table 5 vs. Section 3.1] The reward values in Table 5 (8.59, 15.35, 15.48) are inconsistent with the statement in Section 3.1 that the reward is normalized between [0,1]; either the normalization is not applied in this table or the description is inaccurate.
  3. [Section 4.5] The phrase "SmartGen takes almost 2.31 times more faster than any of the Rankers" is ungrammatical; it should read "is 2.31 times faster."
  4. [Section 4.6] The word "colliqual" appears to be a typo for "colloquial."
  5. [Figure 2] The figure caption does not indicate which test set or dataset the brute-force analysis is performed on, making the motivation difficult to reproduce.
  6. [Section 2.2, Table 1] Table 1 reports a BLEU of 85.83 for the reference repeated K times, but it is unclear which test subset this is computed on and how many sentences it contains; please specify the subset size.

Circularity Check

1 steps flagged · score 4.0 of 10

The reward-as-judge evaluation in Table 5 is partially circular: the RM is trained with top-3 BLEU candidates as preferred responses, while SmartGen is trained on sacreBLEU reward, so the higher reward reported is partly forced by construction.

  1. fitted input called prediction [Sec. 3.2 (Reward Model, Eq. 1); Sec. 4.6 (Reward as judge, Table 5)]
    "Specifically, for each input sample x, we select four preferred responses, denoted by P (including human preference and the top-3 BLEU candidates), with the remaining candidates designated as rejected... Using reward as the judge, we calculated the quality of our translations on Fusion-based systems. From Table 5 we infer that the reward for our approaches SmartGen and SmartGen++ is much higher than LLM-Blender..."

    The RM's preferred set explicitly includes the top-3 BLEU candidates, and the DQN's training reward is the sacreBLEU score of the fusion output ('To train the DQN, we use the sacreBLEU score ... of the final output translation'). Thus the RM used as a judge has been fit to reward the same BLEU-based signal that SmartGen optimizes. Reporting this RM reward as evidence of higher translation quality compares the system against a learned proxy of its own training objective, so the higher reward of SmartGen/SmartGen++ in Table 5 is expected by construction rather than independent validation.

full rationale

The main experimental claims are supported by independent external metrics: Tables 2 and 3 report BLEU, COMET, and chrF++ on Flores, WMT-14, IN22, and private test sets, and the comparison against LLM-Blender, PairRanker, SimCLS, and SummaReranker is not circular. The DQN is legitimately trained with the BLEU of the fusion output as a reward; that is a standard training signal. The one partially circular element is the 'Reward as judge' ablation: the in-house reward model is trained with preferred responses that include top-3 BLEU candidates, so using it to demonstrate quality gains essentially reuses a fitted proxy of the training objective. This does not invalidate the main benchmark tables, but it does make the Table 5 quality claim partially self-referential. The separate concern that the DQN selects top-K candidates by independent Q-values rather than by true group optimization is a modeling and evaluation gap, not a circularity by construction, so it is not scored as a circular step.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claims rest on several domain assumptions about the reliability of BLEU as a reward, the representativeness of the candidate pool, and the effectiveness of LLM-based correction with a reward model trained on BLEU-derived preferences. The DQN's top-K selection also assumes group quality is separable, which is not established and conflicts with the paper's own motivation.

free parameters (2)
  • K (number of selected candidates) = 3
    Fixed number of candidates sent to the fusion block. Chosen in line with Jiang et al. (2023) and not adapted per sentence; the paper lists making K adaptive as a limitation.
  • tau (margin threshold) = 1e-3
    Threshold in the CCB (Algorithm 1) used to decide whether a candidate is corrected. Reported in Table 7, but no sensitivity analysis is provided.
assumptions (5)
  • domain assumption The candidate pool of eight pretrained models is sufficiently diverse and representative for ensembling.
    Used to define the model pool in Section 4.2; no analysis is given for how results depend on pool composition or size.
  • domain assumption The sacreBLEU score of the fusion output is a reliable scalar reward for training the selection policy.
    DQN training uses BLEU of the fused output as reward (Section 3.1); no noise or robustness analysis is presented.
  • domain assumption The reward model trained with preferred responses including top-3 BLEU candidates and human preference generalizes to judge translation quality.
    Used in the CCB (Section 3.2); the training data is private and not released, so generalization cannot be independently assessed.
  • ad hoc to paper An LLM can improve a weak candidate when given the candidate, its reward score, and rejected candidates.
    Core to the CCB (Algorithm 1); the evidence in Table 4 is partially contradictory for Hindi-English and no detailed examples are provided.
  • ad hoc to paper The quality of a candidate group is separable, so selecting top-K individual Q-values approximates the optimal group.
    The DQN selects top-K actions independently (Section 3.1), but the paper's own motivation (Figure 2) shows group optimality is sentence-dependent, implying interactions between candidates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Faster Machine Translation Ensembling with Reinforcement Learning and Competitive Correction." pith.science (2026). https://pith.science/paper/RM2MY7XE

@misc{pith2026250115219,
  author       = {Pith},
  title        = {Pith review of: Faster Machine Translation Ensembling with Reinforcement Learning and Competitive Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RM2MY7XE}},
  note         = {Machine review of arXiv:2501.15219}
}
abstract

Ensembling neural machine translation (NMT) models to produce higher-quality translations than the $L$ individual models has been extensively studied. Recent methods typically employ a candidate selection block (CSB) and an encoder-decoder fusion block (FB), requiring inference across \textit{all} candidate models, leading to significant computational overhead, generally $\Omega(L)$. This paper introduces \textbf{SmartGen}, a reinforcement learning (RL)-based strategy that improves the CSB by selecting a small, fixed number of candidates and identifying optimal groups to pass to the fusion block for each input sentence. Furthermore, previously, the CSB and FB were trained independently, leading to suboptimal NMT performance. Our DQN-based \textbf{SmartGen} addresses this by using feedback from the FB block as a reward during training. We also resolve a key issue in earlier methods, where candidates were passed to the FB without modification, by introducing a Competitive Correction Block (CCB). Finally, we validate our approach with extensive experiments on English-Hindi translation tasks in both directions.

Figures

Figures reproduced from arXiv: 2501.15219 by the authors.

Figure 1
Figure 1. (a)The general strategy of ensembling MT [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. A comparative analysis of the distribution of the number of times candidate triplets are chosen in a) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Detailed block diagram of our proposed methodology. Here DQN+FB represents SmartGen, and [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Demonstration of Inference Time Complex [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: A scatter plot showing the tradeoff of various [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Demonstration of GPT Score vs. Reward Score vs. BLEU on curated samples having idiomatic expressions, cultural nuances and a mixture of colliqual and formal translations. Effect of Competitive Correction Block: In both translation directions, we observed that incorpo￾r…
Figure 7
Figure 7. Figure 7: An example of a detailed prompt that we provide to the CCB block LLM [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. In-Domain African Languages Translation Using LLMs and Multi-armed Bandits

    cs.CL 2025-05 reject novelty 4.0 of 10

    Bandit-based model selection matches or slightly improves on the best single NMT system for in-domain English-to-African translation, but the claimed high-confidence statistical support is absent.

Reference graph

Works this paper leans on

35 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [4]

    Iterative Translation Refinement with Large Language Models

    Iterative translation re- finement with large language models. Preprint, arXiv:2306.03856. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm´an, Edouard Grave, Myle Ott, Luke Zettle- moyer, and Veselin Stoyanov

    work page Pith review arXiv

  2. [6]

    Preprint, arXiv:2010.11125

    Be- yond english-centric multilingual machine transla- tion. Preprint, arXiv:2010.11125. Cong Feng and Jie Zhang

    arXiv 2010

  3. [8]

    arXiv preprint arXiv: 2401.15006

    Airavata: Introducing hindi instruction-tuned llm. arXiv preprint arXiv: 2401.15006. Charles Goddard, Shamane Siriwardhana, Malikeh Ehghaghi, Luke Meyers, Vlad Karpukhin, Brian Benedict, Mark McQuade, and Jacob Solawetz

    arXiv

  4. [9]

    Preprint, arXiv:2403.13257

    Arcee’s mergekit: A toolkit for merging large lan- guage models. Preprint, arXiv:2403.13257. Naman Goyal, Cynthia Gao, Vishrav Chaudhary, Peng- Jen Chen, Guillaume Wenzek, Da Ju, Sanjana Kr- ishnan, Marc’Aurelio Ranzato, Francisco Guzman, and Angela Fan

    arXiv

  5. [11]

    Preprint, arXiv:2402.01680

    Large language model based multi-agents: A survey of progress and challenges. Preprint, arXiv:2402.01680. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun

    arXiv

  6. [13]

    In International Conference on Learning Representations

    Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations. 9 Hieu Hoang, Huda Khayrallah, and Marcin Junczys- Dowmunt. 2024a. On-the-fly fusion of large lan- guage models and machine translation. Preprint, arXiv:2311.08306. Hieu Hoang, Huda Khayrallah, and Marcin Junczys- Dowmunt. 2024b. On-the-fl...

    arXiv 2024

  7. [14]

    A Review of Hybrid and Ensemble in Deep Learning for Natural Language Processing

    A review of hybrid and ensemble in deep learn- ing for natural language processing. Preprint, arXiv:2312.05589. Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin

    work page Pith review arXiv

  8. [15]

    Preprint, arXiv:2407.12391

    Llm inference serving: Sur- vey of recent advances and opportunities. Preprint, arXiv:2407.12391. Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer

    arXiv

Show all 35 references
  1. [16]

    Preprint, arXiv:2001.08210

    Multilingual denoising pre- training for neural machine translation. Preprint, arXiv:2001.08210. Yixin Liu and Pengfei Liu

    2001 arXiv

  2. [17]

    Preprint, arXiv:2106.01890

    Simcls: A simple framework for contrastive learning of abstractive summarization. Preprint, arXiv:2106.01890. Zijun Liu, Yanzhe Zhang, Peng Li, Yang Liu, and Diyi Yang

    arXiv

  3. [19]

    Preprint, arXiv:2311.08692

    Routing to the expert: Efficient reward-guided ensemble of large language models. Preprint, arXiv:2311.08692. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisatt...

    arXiv

  4. [20]

    Preprint, arXiv:2303.17651

    Self-refine: It- erative refinement with self-feedback. Preprint, arXiv:2303.17651. V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller

    arXiv

  5. [22]

    Preprint, arXiv:2303.08774

    Gpt-4 technical report. Preprint, arXiv:2303.08774. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Car- roll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda A...

    arXiv

  6. [23]

    Preprint, arXiv:2203.02155

    Training language models to follow instructions with human feedback. Preprint, arXiv:2203.02155. Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang Wang

    arXiv

  7. [24]

    Preprint, arXiv:2308.03188

    Automatically correcting large language models: Sur- veying the landscape of diverse self-correction strate- gies. Preprint, arXiv:2308.03188. Maja Popovi´c

    arXiv

  8. [25]

    Preprint, arXiv:2305.14878

    Leveraging gpt-4 for automatic translation post-editing. Preprint, arXiv:2305.14878. Mathieu Ravaut, Shafiq Joty, and Nancy F. Chen

    arXiv

  9. [26]

    Preprint, arXiv:2203.06569

    Summareranker: A multi-task mixture-of-experts re- ranking framework for abstractive summarization. Preprint, arXiv:2203.06569. Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie

    arXiv

  10. [27]

    Preprint, arXiv:2211.05100

    Bloom: A 176b-parameter open-access multilingual language model. Preprint, arXiv:2211.05100. 10 Naman Shukla, Arinbj¨orn Kolbeinsson, Lavanya Marla, and Kartik Yellepeddi

    arXiv

  11. [29]

    Preprint, arXiv:2009.01325

    Learn- ing to summarize from human feedback. Preprint, arXiv:2009.01325. Gemma Team

    2009 arXiv

  12. [30]

    Preprint, arXiv:2207.04672

    No language left behind: Scal- ing human-centered machine translation. Preprint, arXiv:2207.04672. Fanqi Wan, Xinting Huang, Deng Cai, Xiaojun Quan, Wei Bi, and Shuming Shi

    arXiv

  13. [31]

    Preprint, arXiv:2401.10491

    Knowl- edge fusion of large language models. Preprint, arXiv:2401.10491. Jiayi Xie, Michael Tashman, John Hoffman, Lee Winikor, and Rouzbeh Gerami

    arXiv

  14. [32]

    Preprint, arXiv:2101.10385

    Online and scalable model selection with multi-armed bandits. Preprint, arXiv:2101.10385. Wenda Xu, Danqing Wang, Liangming Pan, Zhenqiao Song, Markus Freitag, William Yang Wang, and Lei Li

    arXiv

  15. [33]

    Preprint, arXiv:2305.14282

    Instructscore: Explainable text genera- tion evaluation with finegrained feedback. Preprint, arXiv:2305.14282. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel

    arXiv

  16. [34]

    Preprint, arXiv:2010.11934

    mt5: A massively multilin- gual pre-trained text-to-text transformer. Preprint, arXiv:2010.11934. Zhihang Yuan, Yuzhang Shang, Yang Zhou, Zhen Dong, Zhe Zhou, Chenhao Xue, Bingzhe Wu, Zhikai Li, Qingyi Gu, Yong Jae Lee, Yan Yan, Beidi Chen, Guangyu Sun, and Kurt Keutzer

    2010 arXiv

  17. [35]

    Preprint, arXiv:2402.16363

    Llm infer- ence unveiled: Survey and roofline model insights. Preprint, arXiv:2402.16363. A Additional details on datasets and methods A.1 Training and Evaluation Data Statistics Private Data Open-Source Data Training Data 98K - Test Data - IN22-Conv-1.5K - IN22Gen-1.02K - WMT...

    arXiv

  18. [2013]

    Preprint, arXiv:1312.5602

    Playing atari with deep reinforcement learning. Preprint, arXiv:1312.5602. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, and Shyamal Anadkat et al

    arXiv

  19. [2014]

    In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 12–58, Baltimore, Maryland, USA

    Findings of the 2014 workshop on statistical machine translation. In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 12–58, Baltimore, Maryland, USA. Associa- tion for Computational Linguistics. Tianle Cai, Yuhong Li, Zhengyang Geng, Hongwu Peng, Ja...

    2014

  20. [2015]

    Preprint, arXiv:1512.03385

    Deep residual learning for image recogni- tion. Preprint, arXiv:1512.03385. Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen

    arXiv

  21. [2018]

    Preprint, arXiv:1811.01846

    Reinforcement learn- ing based dynamic model selection for short-term load forecasting. Preprint, arXiv:1811.01846. Jay Gala, Pranjal A. Chitale, Raghavan AK, Varun Gumma, Sumanth Doddapaneni, Aswanth Kumar, Janki Nawale, Anupama Sujatha, Ratish Puduppully, Vivek Raghavan, Pra...

    arXiv

  22. [2019]

    Preprint, arXiv:1905.08874

    Adaptive model se- lection framework: An application to airline pricing. Preprint, arXiv:1905.08874. Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea V oss, Alec Radford, Dario Amodei, and Paul Christiano

    1905 arXiv

  23. [2020]

    Preprint, arXiv:1911.02116

    Unsuper- vised cross-lingual representation learning at scale. Preprint, arXiv:1911.02116. Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Man- deep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vi- ta...

    1911 arXiv

  24. [2021]

    Preprint, arXiv:2106.03193

    The flores-101 evaluation benchmark for low-resource and multilingual ma- chine translation. Preprint, arXiv:2106.03193. Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V . Chawla, Olaf Wiest, and Xi- angliang Zhang

    arXiv

  25. [2022]

    Preprint, arXiv:2112.10684

    Efficient large scale lan- guage modeling with mixtures of experts. Preprint, arXiv:2112.10684. Ondˇrej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Johannes Leveling, Christof Monz, Pavel Pecina, Matt Post, Herve Saint- Amand, Radu Soricut, Lucia S...

    arXiv

  26. [2023]

    Preprint, arXiv:2310.02170

    Dynamic llm-agent network: An llm- agent collaboration framework with agent team opti- mization. Preprint, arXiv:2310.02170. Keming Lu, Hongyi Yuan, Runji Lin, Junyang Lin, Zheng Yuan, Chang Zhou, and Jingren Zhou

    arXiv

  27. [2024]

    Preprint, arXiv:2401.10774

    Medusa: Simple llm inference acceleration framework with multiple decoding heads. Preprint, arXiv:2401.10774. Pinzhen Chen, Zhicheng Guo, Barry Haddow, and Kenneth Heafield

    arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.