TALL -- A Trainable Architecture for Enhancing LLM Performance in Low-Resource Languages

Amos Azaria; Moshe Ofer; Orel Zamler

REVIEW 4 major objections 5 minor 20 references

TALL -- A Trainable Architecture for Enhancing LLM Performance in Low-Resource Languages

T0 review · 4 major / 5 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 TALL routes Hebrew through a frozen English LLM and claims more accurate missing-word prediction than direct, fine-tuned, or naive-translation baselines.

desk verdict TALL's architecture is genuinely new, but the paper never clarifies whether the Hebrew encoder sees the missing word, so the headline gain could be answer leakage rather than cross-lingual transfer. read the letter →

arxiv 2506.05057 v1 pith:57WWHJAJ submitted 2025-06-05 cs.CL cs.AI

Moshe Ofer , Orel Zamler , Amos Azaria This is my paper

classification cs.CLcs.AI

keywords low-resourcelanguagescross-lingualtransferparameter-efficientadaptationfrozenLLMtranslationmodelsHebrewwordpredictiondimensionalignmentadaptersTALL

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

The paper tries to establish that a frozen English-centred LLM can be made useful for a low-resource language by wrapping it in trainable translation bridges rather than retraining it. On the task of predicting the final word of a Hebrew sentence, TALL reports 5.59% accuracy with bloomz-560m and 5.15% with QWEN2.5-0.5b on WIKI-SLVM, while the next-best baselines reach 2.93% and 3.99%. Because only about 14% of the parameters are trainable, the claimed gain is attributed to the alignment and cross-attention modules, not to expensive fine-tuning of the large models. A reader should care because the recipe, if it holds, is a lightweight route to extending strong English LLMs to languages they were barely trained on.

What carries the argument

The load-bearing mechanism is the trainable cross-lingual bridge between frozen translation models and a frozen LLM. A Marian Hebrew-to-English encoder produces hidden states for the truncated Hebrew sentence; a dimension-alignment adapter (a fully connected layer with layer normalization and GELU activation) projects those states into the LLM's embedding dimension; a custom transformer conditions on the English translation tokens and cross-attends to the aligned Hebrew states; and the frozen English LLM performs next-token prediction. A second alignment adapter and custom encoder then adapt the LLM's hidden states for a frozen English-to-Hebrew Marian decoder, which scores the candidate Hebrew final word. Training uses teacher forcing with a loss confined to the final token position, because every earlier token is already supplied by the translation, and the two custom transformers are derived from Marian configurations.

What would settle it

Build a Hebrew test set in which the missing final word is a verb and the English translation moves the counterpart of that verb away from the final position; if TALL's accuracy collapses on those sentences while baselines hold steady, the assumed alignment between the missing Hebrew word and the English next token is the load-bearing component.

Watch

Extended reading notes

Core claim

Stated on the paper's own terms, TALL is a seven-stage pipeline: a frozen Hebrew-to-English encoder, a dimension-alignment adapter, a custom transformer that cross-attends to the aligned Hebrew states, the frozen English LLM, a second alignment adapter, a second custom transformer, and a frozen English-to-Hebrew decoder. The central discovery claim is that this pipeline, trained only on the final missing token of 256,000 Hebrew news sentences, consistently outperforms direct Hebrew prompting, Hebrew fine-tuning, naive translate-to-English-and-back, soft prompting, and a from-scratch language model on both WIKI-SLVM and HE-LYRICS. The reported headline numbers are 5.59% versus 2.93% for bloomz-560m and 5.15% versus 3.99% for QWEN2.5-0.5b on the Wikipedia-style test set. The authors take this as evidence that the linguistic knowledge stored in a high-resource LLM becomes reachable for a low-resource language once trainable bridges connect the two representation spaces.

Load-bearing premise

The load-bearing premise is that translating a Hebrew sentence without its last word into English and asking a frozen English LLM to predict the next token yields a reliable, positionally aligned signal for the missing Hebrew word, despite Hebrew and English ordering words differently.

Editorial extensions

If this is right

A frozen LLM with little or no exposure to a target language can outperform direct use of that LLM in the target language, provided bilingual translation models exist.
The same architecture should transfer to another low-resource language by swapping the two Marian translation models, since the rest of the pipeline is language-agnostic.
The reported gains hold on two different domains, so the trained bridge is not limited to the news domain it was trained on.
The current supervision scheme only optimizes the last token, so the method is directly validated for sentence-final word prediction rather than general language modelling.

Reading between the lines

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

The paper never verifies that the English token predicted by the frozen LLM corresponds positionally to the missing Hebrew word, so the alignment step is an unmeasured assumption inside the pipeline.
At 5.59% accuracy, the task remains very hard in absolute terms; the architecture improves a difficult guessing task and has not yet been shown to help full-sentence Hebrew generation.
A stricter version of the same idea would mask a mid-sentence word; TALL's final-token-only training signal would need to change before that test is meaningful.
The method inherits the quality of the language pair's translation models, so for extremely low-resource languages without adequate MT, this particular bridge cannot be built as described.

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.

Desk Editor's Note

read the letter

Colleague,

The short version: the architecture is a real combo — two frozen Marian translation models, a frozen LLM, and trainable cross-attention transformers with dimension alignment — and the code is public. But the paper's central result may be an artifact of an underspecified input pipeline. Section 3.1 says the translated sentence is truncated, but it never says whether the Hebrew encoder input is also truncated. If the encoder sees the full sentence, the custom transformer can copy the missing word from cross-attention, and the 5.59% vs 2.93% gap is formatting, not transfer. This is the first thing to check in the code.

The good: the combination is not in CoTR, TaCo, or TALENT, which only use translation in prompting. The final-token-only objective is unusual and the limitations section honestly flags it. The paper reports two dataset-domain evaluations and gives detailed parameter tables.

The soft spots: no error bars or significance tests anywhere, so the gains could be noise. The fine-tuned baseline dropping below direct Hebrew (0.16 vs 0.63 on WIKI-SLVM with bloomz) is suspicious and unexplained; it might be a real domain mismatch, but it needs a sentence. 'Parameter-efficient' is doing heavy lifting: 14% trainable is not lightweight, and the trainable custom transformer is over 100M parameters. The abstract's 'enhancing LLM performance in low-resource languages' overclaims a task that is single final-token prediction on one language.

The leakage ambiguity is the load-bearing issue. The stage 3 parenthetical restricts truncation to the translated sentence; it does not restrict the source encoder. The appendix is consistent with both readings. If the source encoder receives the complete sentence, the cross-attention path gives the model the answer. That is not a minor methodological quibble.

Who this is for: people working on cross-lingual transfer with translation models, and anyone interested in leakage artifacts in multi-stage architectures. I would not cite it until the input format is clarified and the numbers are rerun with error bars. It still deserves a serious referee — the architecture is novel and the code is available, so the decisive test is cheap. Send it to review, expect major revision or rejection depending on what the code shows.

Referee Report

4 major / 5 minor

Summary. TALL is a modular architecture that combines a frozen English-centric LLM with two frozen Marian translation models and a small number of trainable adapters and custom transformers. The paper's central claim is that encoding a Hebrew sentence prefix, translating it into English, and feeding the resulting representation through the frozen LLM allows the system to predict the missing final Hebrew token more accurately than direct Hebrew prediction, naive translation, soft prompting, or fine-tuning. Experiments on two Hebrew datasets (WIKI-SLVM and HE-LYRICS) with bloomz-560m and QWEN2.5-0.5b report accuracies around 5%, consistently above baselines. The manuscript also emphasizes parameter efficiency, reporting that roughly 86% of all parameters are frozen.

Significance. If the empirical results were reliable, TALL would be a plausible parameter-efficient recipe for transferring high-resource LLM knowledge to a morphologically rich low-resource language. The manuscript is commendably explicit about module-level parameter counts, releases code, and candidly lists limitations such as inference cost and reliance on machine translation quality. These are genuine strengths. However, the significance is currently limited by three factors: the evaluation is restricted to a single final-token prediction task, the experimental section lacks repeated runs and statistical tests, and there is an unresolved ambiguity about whether the source-language encoder sees the answer token during training and inference. The contribution is therefore promising but not yet established at the level claimed.

major comments (4)

[Section 3.1, stages 1 and 3] The manuscript never states whether the source-language encoder in stage 1 receives the full Hebrew sentence or the same truncated sentence used for translation. The parenthetical in stage 3 says that during training 'the last word is removed before the sentence is translated,' but this does not specify the input to the LR-HR encoder. If the source encoder receives the full sentence, the aligned hidden states available to the Custom LR-HR Transformer through cross-attention contain the final Hebrew word, allowing the model to copy the answer directly from the source rather than leveraging the English LLM. Because this concerns the core mechanism of the reported gains, please specify the exact input format for the source encoder at both training and inference time, and add an ablation in which the source input is truncated identically to the translation input.
[Section 4.2, Table 1] Table 1 reports a single run per condition, with no standard deviation, confidence interval, or significance test. Several headline differences are small; for example, on WIKI-SLVM with QWEN2.5-0.5b, TALL achieves 5.15% versus 3.99% for the next best method. Without repeated runs, these numbers cannot be distinguished from sampling noise, especially given the low absolute accuracies. Please report multiple seeds with means and standard deviations, and use an appropriate paired significance test over the same evaluation items.
[Section 4.1 and Table 1] The fine-tuned baseline is surprisingly worse than direct Hebrew prediction: for bloomz-560m on WIKI-SLVM, fine-tuning yields 0.16% accuracy while direct Hebrew yields 0.63%, and the same pattern appears on HE-LYRICS. This is not explained and suggests a possible training or evaluation mismatch in the fine-tuning pipeline, such as a tokenization issue or a mismatch in how the final token is scored. Please verify the fine-tuning baseline, report the exact fine-tuning configuration, and clarify why performance degrades so sharply.
[Section 3.2 and Section 6] The architecture is trained only on the final token of the target sequence, and the evaluation is exactly that final-token task. The abstract and conclusion claim more broadly that TALL 'enhances LLM performance in low-resource languages.' As presented, the evidence supports only a narrow word-prediction task; the broader claim is not tested. Please either qualify the claims or add experiments on a more general generation task to support the stated scope.

minor comments (5)

[Section 3.2] The training data is described as '256,000 Hebrew sentences from news articles,' but the exact corpus and preprocessing script are not named; please state the source and confirm that the news training set does not overlap with WIKI-SLVM or HE-LYRICS evaluation sentences.
[Section 4.1] The 'Fine-tuned' baseline is inconsistently described as fine-tuning in the main text and as 'parameter-efficient fine-tuning' in the appendix; please specify whether adapters or LoRA were used and exactly which parameters were updated.
[Figure 1] The text in Figure 1 is very small in the rendered version; please provide a higher-resolution figure so that the stage labels and data-flow arrows are legible.
[Tables 3 and 4] The autoencoder dimensions are listed in an inconsistent format; for example, Table 3 lists '1024→2048→1024' while Table 4 lists '1024→1792, 1792→896,' which makes the layer sizes unclear. Please use a uniform notation for the MLP shapes.
[Appendix: Evaluation Framework] The evaluation uses temperature 0.7 with top-k 50 and top-p 0.95 sampling; for a word-accuracy metric, greedy decoding or majority voting over multiple samples is more standard. Please report the number of samples and the sensitivity of the results to these decoding choices.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TALL's reported gains are empirical held-out evaluations, not derived from its inputs by construction.

full rationale

The paper's central claim is an empirical comparison: TALL is trained on 256,000 news sentences and evaluated on held-out test sets (WIKI-SLVM and HE-LYRICS) against several baselines. No trainable parameter is fitted to the test data, and no test-set quantity is renamed as a prediction. The final-token-only loss is an explicitly acknowledged design constraint of the architecture (Section 6, 'Single-Token Learning Signal'), not a hidden circular loop: the model is trained and evaluated on the same task definition, which aligns the design with the metric but does not make the result true by definition. There is no load-bearing self-citation: the cited prior work (Marian, adapters, translation prompting, etc.) is external and is not used to forbid alternatives or to import an unverified premise. The only potentially serious issue is the formatting ambiguity in Section 3.1, where the parenthetical 'during training, the last word is removed before the sentence is translated' could leave open whether the source-language encoder also receives the full sentence including the missing final word; if it does, cross-attention could copy the answer from source hidden states. That would be an evaluation-validity or data-leakage problem, not circularity in the sense of a claim reducing by construction to its inputs, and the paper's Section 6 limitation on error propagation and the single-token learning signal are honest statements rather than circular moves. The derivation chain is experimental and externally benchmarked, so the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 3 invented entities

The central claim depends on a large number of trainable parameters in custom modules and on the quality of frozen pre-trained components. No explicit hand-fitted constants are used, but the architecture introduces several new trainable modules whose behavior is only evaluated on one task and one language.

free parameters (2)

inference sampling parameters = T=0.7, top-k=50, top-p=0.95
Chosen by hand for evaluation; affects reported accuracy.
TALL training hyperparameters = not specified
The paper states AdamW with cosine annealing but does not report initial learning rate, epochs, or batch size, all of which affect final performance.

assumptions (3)

domain assumption Pre-trained models (Marian MT, bloomz-560m, QWEN2.5-0.5b) provide useful representations for Hebrew-English translation and English language modeling.
The architecture depends on the quality of these frozen components. Any bias or error in them will propagate.
domain assumption The news dataset of 256,000 Hebrew sentences is sufficient to train the trainable components.
The paper does not provide analysis of whether this data size is sufficient.
domain assumption Per-token accuracy on the last-word prediction is an appropriate metric for evaluating low-resource language performance.
The paper uses this metric and does not validate its correlation with other language understanding measures.

invented entities (3)

Dimension alignment adapters
purpose: Project hidden states between the encoder, LLM, and decoder dimensional spaces.
Trainable MLP modules introduced by the paper; no external evidence beyond the reported experiments.
Custom LR-HR transformer
purpose: Cross-attend to aligned low-resource encoder states while processing high-resource token embeddings.
A trainable transformer module without independent validation outside the paper.
Custom HR-LR encoder
purpose: Refine representations before final decoding in the target language.
A trainable transformer module introduced for the architecture.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TALL -- A Trainable Architecture for Enhancing LLM Performance in Low-Resource Languages." pith.science (2026). https://pith.science/paper/57WWHJAJ

@misc{pith2026250605057,
  author       = {Pith},
  title        = {Pith review of: TALL -- A Trainable Architecture for Enhancing LLM Performance in Low-Resource Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/57WWHJAJ}},
  note         = {Machine review of arXiv:2506.05057}
}

read the original abstract

Large Language Models (LLMs) excel in high-resource languages but struggle with low-resource languages due to limited training data. This paper presents TALL (Trainable Architecture for Enhancing LLM Performance in Low-Resource Languages), which integrates an LLM with two bilingual translation models. TALL transforms low-resource inputs into high-resource representations, leveraging the LLM's capabilities while preserving linguistic features through dimension alignment layers and custom transformers. Our experiments on Hebrew demonstrate significant improvements over several baselines, including direct use, naive translation, and fine-tuning approaches. The architecture employs a parameter-efficient strategy, freezing pre-trained components while training only lightweight adapter modules, balancing computational efficiency with performance gains.

Figures

Figures reproduced from arXiv: 2506.05057 by the authors.

**Figure 1.** Overview of the TALL architecture with numbered components (1-7) corresponding to the processing stages described in Section 3.1. * For input high-resource sentences, the last word is removed before translation during training; ** For target sentences, teacher forcing is used during training while outputs are generated auto-regressively during inference. 5. Dimension Alignment (LLM to Target): A second alignment a… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 4 canonical work pages

[1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Samuel Cahyawijaya, Holy Lovenia, and Pascale Fung. 2024. https://doi.org/10.18653/V1/2024.NAACL-LONG.24 Llms are few-shot in-context low-resource language learners . arXiv preprint arXiv:2403.16512, pages 405--433

arXiv 2024
[4]

Tejas Deshpande, Nidhi Kowtal, and Raviraj Joshi. 2024. https://arxiv.org/abs/2409.04512 Chain-of-translation prompting (cotr): A novel prompting technique for low resource languages . Preprint, arXiv:2409.04512

arXiv 2024
[5]

Mozhdeh Gheini, Xiang Ren, and Jonathan May. 2021. https://arxiv.org/abs/2104.08771 On the strengths of cross-attention in pretrained transformers for machine translation . CoRR, abs/2104.08771

arXiv 2021
[6]

Ping Guo, Yubing Ren, Yue Hu, Yunpeng Li, Jiarui Zhang, Xingsheng Zhang, and Heyan Huang. 2024. https://aclanthology.org/2024.lrec-main.1362/ Teaching large language models to translate on low-resource languages with textbook prompting . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluati...

work page 2024
[7]

Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. https://arxiv.org/abs/1902.00751 Parameter-efficient transfer learning for NLP . CoRR, abs/1902.00751

arXiv 2019
[8]

Marcin Junczys-Dowmunt, Alexandra Birch, Roman Grundkiewicz, Kenneth Heafield, Tomasz Dwojak, Hieu Hoang, Tom Neckermann, Frank Seide, Ulrich Germann, Alham Fikri Aji, Nikolay Bogoychev, Andr\'e Martins, and Alexandra F. Birch. 2018. https://doi.org/10.18653/v1/P18-4020 Marian: Fast neural machine translation in c++ . In Proceedings of ACL 2018: System De...

doi:10.18653/v1/p18-4020 2018

Show all 20 references

[9]

LCM-team, Lo \" c Barrault, Paul - Ambroise Duquenne, Maha Elbayad, Artyom Kozhevnikov, Belen Alastruey, Pierre Andrews, Mariano Coria, Guillaume Couairon, Marta R. Costa - juss \` a , David Dale, Hady Elsahar, Kevin Heffernan, Jo \ a o Maria Janeiro, Tuan Tran, Christophe Rop...

2024 doi arXiv
[10]

Brian Lester, Rami Al - Rfou, and Noah Constant. 2021. https://doi.org/10.18653/V1/2021.EMNLP-MAIN.243 The power of scale for parameter-efficient prompt tuning . pages 3045--3059

2021 doi
[11]

Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M. Saiful Bari, Sheng Shen, Zheng Xin Yong, Hailey Schoelkopf, Xiangru Tang, Dragomir Radev, Alham Fikri Aji, Khalid Almubarak, Samuel Albanie, Zaid Alyafeai, Albert Webson, Edward...

2023 doi
[12]

NLPH. 2019. Svlm-hebrew-wikipedia-corpus. https://github.com/NLPH/SVLM-Hebrew-Wikipedia-Corpus

2019
[13]

Norod78. 2023. Hebrewstageandlyricswithnewlines. https://huggingface.co/datasets/Norod78/HebrewStageAndLyricsWithNewLines

2023
[14]

Sinno Jialin Pan and Qiang Yang. 2010. https://doi.org/10.1109/TKDE.2009.191 A survey on transfer learning . IEEE Transactions on Knowledge and Data Engineering, 22(10):1345--1359

2010 doi
[15]

Jonas Pfeiffer, Aishwarya Kamath, Andreas R \" u ckl \' e , Kyunghyun Cho, and Iryna Gurevych. 2020 a . https://arxiv.org/abs/2005.00247 Adapterfusion: Non-destructive task composition for transfer learning . CoRR, abs/2005.00247

2020 arXiv
[16]

Jonas Pfeiffer, Ivan Vuli \'c , Iryna Gurevych, and Sebastian Ruder. 2020 b . https://doi.org/10.18653/v1/2020.emnlp-main.617 MAD-X : A n A dapter- B ased F ramework for M ulti- T ask C ross- L ingual T ransfer . In Proceedings of the 2020 Conference on Empirical Methods in Na...

2020 doi
[17]

Qwen-Team. 2024. https://qwenlm.github.io/blog/qwen2.5/ Qwen2.5: A party of foundation models

2024
[18]

Sylvestre - Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. 2017. https://arxiv.org/abs/1705.08045 Learning multiple visual domains with residual adapters . CoRR, abs/1705.08045

2017 arXiv
[19]

Bibek Upadhayay and Vahid Behzadan. 2024. https://openreview.net/forum?id=02MLWBj8HP Taco: Enhancing cross-lingual transfer for low-resource languages in LLM s through translation-assisted chain-of-thought processes . In 5th Workshop on practical ML for limited/low resource settings

2024
[20]

Gomez, Lukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. https://arxiv.org/abs/1706.03762 Attention is all you need . CoRR, abs/1706.03762

2017 arXiv

Pith tools

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

[1] [1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Samuel Cahyawijaya, Holy Lovenia, and Pascale Fung. 2024. https://doi.org/10.18653/V1/2024.NAACL-LONG.24 Llms are few-shot in-context low-resource language learners . arXiv preprint arXiv:2403.16512, pages 405--433

arXiv 2024

[4] [4]

Tejas Deshpande, Nidhi Kowtal, and Raviraj Joshi. 2024. https://arxiv.org/abs/2409.04512 Chain-of-translation prompting (cotr): A novel prompting technique for low resource languages . Preprint, arXiv:2409.04512

arXiv 2024

[5] [5]

Mozhdeh Gheini, Xiang Ren, and Jonathan May. 2021. https://arxiv.org/abs/2104.08771 On the strengths of cross-attention in pretrained transformers for machine translation . CoRR, abs/2104.08771

arXiv 2021

[6] [6]

Ping Guo, Yubing Ren, Yue Hu, Yunpeng Li, Jiarui Zhang, Xingsheng Zhang, and Heyan Huang. 2024. https://aclanthology.org/2024.lrec-main.1362/ Teaching large language models to translate on low-resource languages with textbook prompting . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluati...

work page 2024

[7] [7]

Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. https://arxiv.org/abs/1902.00751 Parameter-efficient transfer learning for NLP . CoRR, abs/1902.00751

arXiv 2019

[8] [8]

Marcin Junczys-Dowmunt, Alexandra Birch, Roman Grundkiewicz, Kenneth Heafield, Tomasz Dwojak, Hieu Hoang, Tom Neckermann, Frank Seide, Ulrich Germann, Alham Fikri Aji, Nikolay Bogoychev, Andr\'e Martins, and Alexandra F. Birch. 2018. https://doi.org/10.18653/v1/P18-4020 Marian: Fast neural machine translation in c++ . In Proceedings of ACL 2018: System De...

doi:10.18653/v1/p18-4020 2018