Pith. sign in

REVIEW 4 major objections 6 minor 25 references

Homeostasis and Sparsity in Transformer

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper argues that adding two homeostatic sparsity mechanisms to the transformer's self-attention and block outputs improves generalization in machine translation, with the best model reaching 0.3062 BLEU on Multi30K versus 0.2768 for…

desk verdict Novel homeostatic sparsity mechanisms, but the reported BLEU gains are single-run best-of-grid results that do not support the claim of significant improvement. read the letter →

arxiv 2412.00503 v3 pith:HKCHWY3Y submitted 2024-11-30 cs.LG cs.AI

Leonid Kotyuzanskiy , Artem Klimov This is my paper · ORCID
classification cs.LGcs.AI
keywords transformersparsedistributedrepresentationshomeostasisnaturallanguageprocessingmachinetranslationkWTAself-attentionMulti30K
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 aims to establish that the classical transformer's generalization can be improved by adding homeostatic sparsity mechanisms inspired by neocortical models. The first mechanism, RFB-kWTA, keeps a running cache of per-head activation statistics and boosts rarely activated units before applying k-winners-take-all sparsification; the second, 'Smart' Inhibition, stochastically masks activations with probabilities set by the same statistics. On the Multi30K English-German dataset, the authors report that the best combination reaches 0.3062 BLEU, above the 0.2768 BLEU of the classical transformer, and that a small model with RFB-kWTA reaches 0.3025 BLEU, above the 0.2751 BLEU of a ten-times-larger classical model. The paper also reports that plain kWTA improves training-data memorization while hurting generalization, and that the homeostatic boost is what restores it. If true, the result suggests that activation-statistics-based self-regulation can substitute for added parameters.

What carries the argument

The central object is the activation-statistics cache used by both mechanisms: a FIFO tensor that counts, over the last Q training steps, how often each embedding position survived the kWTA mask, with shape $(H, Q, D_h)$ inside self-attention and $(Q, D_h)$ at the transformer-block output. RFB-kWTA is k-winners-take-all sparsification preceded by a boosting step: activations are scaled so that historically rare units are amplified before the top-k mask is chosen. 'Smart' Inhibition converts the same statistics into per-unit sampling probabilities, clips them to the interval [0.01, 0.99], and adjusts the mask so the realized sparsity stays near the target $s$. These two operations carry the argument because they allow low-magnitude features to survive the sparsity mask periodically, which the paper identifies as the source of improved generalization.

What would settle it

Retrain the classical transformer on Multi30K with the paper's stated parameters and optimizer (Adam with learning rate $10^{{-4}}$, 181k steps) and check whether it reproduces the reported 0.2768 BLEU; if a properly tuned baseline reaches or surpasses 0.3062 BLEU, the claimed improvement disappears.

Watch

Extended reading notes

Core claim

On the paper's own terms: sparsifying the self-attention output with kWTA removes the smallest features, which speeds memorization but damages generalization because those features encode 'subtle' patterns. RFB-kWTA repairs this by accumulating, in a FIFO cache of shape $(H, Q, D_h)$, the activation history of each head, then amplifying each activation in inverse proportion to its historical frequency before the top-k mask is computed. 'Smart' Inhibition converts the same statistics into Bernoulli sampling probabilities, keeps the expected sparsity fixed through a median-shift correction, and can replace dropout at the block output. The authors claim that this combination yields 0.3062 BLEU on Multi30K, that every model containing a homeostatic mechanism beats the classical transformer, and that the small RFB-kWTA model outperforms the big classical transformer despite having roughly ten times fewer parameters. The conclusion is that self-regulated strengthening of rare features, not sparsity alone, is what drives the generalization gain.

Load-bearing premise

The load-bearing premise is that the reimplemented classical transformer is a faithful, competitive baseline; if that baseline is undertuned or its training setup is misreported, the BLEU advantages attributed to homeostatic sparsity could be artifacts of comparison.

Editorial extensions

If this is right

  • Models that use plain kWTA in self-attention memorize the training set faster but generalize worse than the classical transformer, establishing a sparsity-driven memorization-generalization tradeoff.
  • A small model with RFB-kWTA at s=0.8 reaches 0.3025 BLEU, beating the big classical transformer's 0.2751 BLEU despite having about ten times fewer parameters.
  • Replacing block-output dropout with 'Smart' Inhibition improves results, and combining it with RFB-kWTA in attention gives the best overall score of 0.3062 BLEU.
  • Across every sparsity coefficient tested, every model containing a homeostatic mechanism outperforms the classical transformer on Multi30K.
  • The authors hypothesize that the mechanisms will transfer to other attention-based tasks such as image processing, while noting this remains to be tested.

Reading between the lines

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

  • Beyond the paper: the paper does not report the computational or memory overhead of maintaining the activation-statistics cache, so an independent cost-benefit measurement would clarify whether the BLEU gains justify the added machinery in production.
  • Beyond the paper: the paper leaves the dependence on cache size QAtt unexplored; varying Q while holding s fixed would show whether the benefit comes from long-run statistics or simply from temporal smoothing of the mask.
  • Beyond the paper: sharing one activation statistics cache across attention heads, which the paper lists as a future experiment, would test whether per-head statistics are load-bearing or redundant.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 manuscript introduces two mechanisms for introducing sparsity and homeostasis into Transformer layers: Rare Features Boosting kWTA (RFB-kWTA), which applies kWTA after boosting rarely activated features using running activation statistics, and "Smart" Inhibition, which samples binary masks from a probability vector derived from those statistics. The mechanisms are inserted at the self-attention output and at the output of the transformer block, and are evaluated on English-to-German Multi30K translation with small, base, and big Transformers. The paper reports BLEU gains over a classical Transformer and over a dropout-only control, culminating in 0.3062 BLEU for the base model with Smart Inhibition at both positions.

Significance. The underlying hypothesis — that temporal statistics of activations can be used as a homeostatic regularizer that selects which features survive sparsification — is interesting and connects to a productive line of biologically inspired sparsity research. The paper deserves credit for testing several insertion points and cache sizes and for including a dropout-only control. However, the empirical evidence is currently not sufficient: all comparisons are single-run, the headline number is selected from a hyperparameter grid, the classical baseline is not externally validated, and the memorization metric is misdefined. If the missing evidence were supplied, the paper could be a useful empirical contribution; as it stands, its central claim is not established.

major comments (4)
  1. [§3.4, Table 3] The reported improvement is not statistically supported. The best result (Base, Smart Inhibition at both positions, s=0.9, QBO=16, BLEU 0.3062) is a single run, and it is the maximum over the QBO grid {1, 16, 256, 1024} and the chosen sparsity settings; the table also reports a 0.3025 result from a grid over s. The gap over the dropout-only control (0.3007) is 0.0055 BLEU, which on a 1,024-sentence test set is likely within run-to-run noise. The word "significantly" therefore has no statistical meaning in this paper. At minimum, report multiple seeds per configuration, error bars, and a paired significance test, and either fix the hyperparameters before looking at test BLEU or explicitly account for selection.
  2. [§3.1, §2.6, §3.2] The classical Transformer baseline is unvalidated and the training protocol is internally inconsistent. Section 3.1 states that the models are "reproduced from work [1] with the same parameters", but no published Multi30K number is cited for comparison, no baseline training curves are shown, and Section 2.6 says the models were trained for 181k steps while Section 3.2 says the comparative study was run for 45k steps. Without a verified baseline, the larger apparent gains over the classical Transformer (for example, 0.3025 versus 0.2751 in the conclusion) cannot be distinguished from an undertrained comparison model. Please provide the full protocol and validate the baseline externally.
  3. [§2.5, Eq. (9)] The IMI definition does not measure the quantity it is claimed to measure. Equation (9) is described as "the ratio of the area under the curve of the metric trained by the model to the area under the curve of the ideal model", but the sum telescopes to (f_E − f_1)/(2(E−1)), i.e., a scaled difference of the first and last epoch values. Consequently, the IMI column in Table 2 and the discussion of memorization speed are based on a misdefined quantity; please redefine the metric or remove the area-ratio interpretation.
  4. [§2.3–§2.4, Eqs. (6)–(8)] The method has several hand-picked constants whose sensitivity is not examined. These include a=0.99, b=0.01, the exponent 0.83 in Eq. (7), the threshold δ in Eq. (8), and the cache sizes QAtt and QBO. Because the best configuration is selected over part of this set, it is not clear whether the favorable BLEU results are due to the homeostatic mechanism itself or to tuning of these constants. A sensitivity analysis, or a fixed a priori choice of the constants, is required to support the causal interpretation.
minor comments (6)
  1. [§2.3, Eq. (5)] The indices in the sum are confusing: the tensor T is indexed by (H, Q, Dh), but the right-hand side of Eq. (5) reads "thi,p,j"; please make the summation index and the component index consistent.
  2. [§2.4, Eq. (8)] The threshold δ is introduced but never defined or assigned a value in the experiments.
  3. [§3.1] The small model is said to be "reproduced from work [1] with the same parameters", but a small model is not present in [1]; clarify which hyperparameters were changed.
  4. [§3.3, Figures 3–4] Several claims of "significant" improvement are based on visual inspection of single training curves and lack error bars or statistical tests.
  5. [Throughout] There are numerous typographical issues: "homeostazis" in the title and abstract, "Sof tmax" in Eq. (2), "thinning" for sparsification, and "interblock" for the block output position; please proofread carefully.
  6. [Table 3] The notation QAtt and QBO is not defined in the table caption; state that QAtt is fixed at 256 and QBO is the cache size at the block output, and explain why QAtt is not varied.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the test-set hyperparameter sweep is a statistical concern, not a circular step.

full rationale

The paper's proposed mechanisms are defined by explicit formulas (Eqs. 1–8), and the experimental claims compare models trained on Multi30K against a classical transformer baseline reproduced from Vaswani et al. There are no author self-citations invoked as load-bearing evidence, no uniqueness theorem imported from prior work by the same authors, and no ansatz smuggled in via citation. The measured BLEU differences are empirical outcomes, not quantities constructed to equal their inputs. The headline 0.3062 in Table 3 is the best score among a transparently tabulated grid of s and QBO values, so it is a selected maximum rather than an a priori prediction; this is a legitimate statistical-reporting concern, but it is not a circular reduction because the paper does not present the swept configurations as a prediction derived from the mechanism. Other weaknesses—the unvalidated classical-transformer baseline, the inconsistent step counts (181k in Section 2.6 vs. 45k in Section 3.2), and the acknowledged lack of a QAtt-dependence study in Section 4—are correctness and reproducibility risks, not instances of the derivation chain reducing to its own inputs. No circular step can be exhibited from the paper's own equations or citations.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities; its mechanisms are built from hand-chosen constants and cache designs. The central claim depends on six free parameters (s, QAtt, QBO, a, b, and the 0.83 exponent, plus an unreported delta), several of which are tuned on the validation/test performance, and on assumptions about the reliability of the activation statistics cache.

free parameters (6)
  • sparsity coefficient s = 0.2, 0.5, 0.7, 0.8, 0.9
    Controls the fraction of kept activations in kWTA and Smart Inhibition; swept across runs and the best value (0.8 or 0.9) is reported.
  • attention cache size QAtt = 256
    FIFO cache length for activation statistics in attention; fixed at 256 without reported sensitivity analysis.
  • block-output cache size QBO = 1, 4, 16, 256, 512, 1024 (best: 16)
    FIFO cache length for Smart Inhibition at block output; swept and the value maximizing BLEU is selected.
  • max/min sampling probabilities a, b = a=0.99, b=0.01
    Hand-chosen bounds in Eq. 7; no sensitivity analysis.
  • probability exponent 0.83 = 0.83
    Hand-chosen exponent in Eq. 7 mapping statistics to sampling probabilities.
  • recalibration threshold delta = not reported
    Threshold in Eq. 8 for adjusting probabilities to maintain sparsity; value is never given.
assumptions (3)
  • domain assumption The transformer architecture of Vaswani et al. is taken as given and reimplemented faithfully.
    The paper builds on [1] and assumes the reimplementation matches the original model behavior.
  • domain assumption Running activation statistics from a FIFO cache of size Q are a representative estimate of activation rarity over training.
    The boosting and sampling mechanisms (Eqs. 4-7) rely on this estimator, but no analysis is given for cache size sufficiency or stationarity.
  • domain assumption The enhancement factor in Eq. 6 preserves the ordering of activations well enough that kWTA still selects the intended top-k features.
    The formula divides by the k-th order statistic v; whether this keeps the top-k set stable across training is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Homeostasis and Sparsity in Transformer." pith.science (2026). https://pith.science/paper/HKCHWY3Y

@misc{pith2026241200503,
  author       = {Pith},
  title        = {Pith review of: Homeostasis and Sparsity in Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HKCHWY3Y}},
  note         = {Machine review of arXiv:2412.00503}
}
read the original abstract

The transformer architecture has become an integral part of the field of modern neural networks, playing a crucial role in a variety of tasks, such as text generation, machine translation, image and audio processing, among others. There is also an alternative approach to building intelligent systems, proposed by Jeff Hawkins and inspired by the processes occurring in the neocortex. In our article we want to combine some of these ideas and to propose the use of homeostasis mechanisms, such as RFB-kWTA and "Smart" Inhibition, in the attention mechanism of the transformer and at the output of the transformer block, as well as conducting an experiment involving the introduction of sparse distributed representations of the transformer at various points. RFB-kWTA utilizes statistics of layer activations across time to adjust the entire layer, enhancing the values of rare activations while reducing those of frequent ones. "Smart" Inhibition also uses activation statistics to sample sparsity masks, with rarer activation times are more likely to be activated. Our proposed mechanisms significantly outperform the classical transformer 0.2768 BLEU and a model that only makes use of dropout in the attention mechanism and output of the transformer block 0.3007 BLEU, achieving a score of 0.3062 on the Multi30K dataset.

Figures

Figures reproduced from arXiv: 2412.00503 by the authors.

Figure 1
Figure 1. BLEU graph based on training data for small, small-kWTA [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. BLEU graph based on validation data for models small, small with kWTA [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. BLEU graph based on validation data for models small, small RFB-kWTA [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: BLEU graph based on validation data for models small, base, big and small RFB-kWTA [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 7 canonical work pages

  1. [1]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 6000–6010, Red Hook, NY , USA, 2017. Curran Associates Inc. ISBN 9781510860964

    2017

  2. [2]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b, 2023. URL https://arxi...

    arXiv 2023

  3. [3]

    Llama: Open and efficient foundation language models, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. URL https://arxiv.org/abs/2302.13971

    arXiv 2023

  4. [4]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Anto...

    arXiv 2024

  5. [5]

    Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, Alon Benhaim, Misha Bilenko, Johan Bjorck, Sébastien Bubeck, Martin Cai, Qin Cai, Vishrav Chaudhary, Dong Chen, Dongdong Chen, Weizhu Chen, Yen-Chun Chen, Yi-Ling Chen, Hao Cheng, Parul Chopra, Xiyang Dai, Matt...

    arXiv 2024

  6. [6]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. URL https://arxiv.org/abs/2010.11929

    arXiv 2021

  7. [7]

    Audio transformers:transformer architectures for large scale audio understanding

    Prateek Verma and Jonathan Berger. Audio transformers:transformer architectures for large scale audio understanding. adieu convolutions, 2021. URL https://arxiv.org/abs/2105.00335

    arXiv 2021

  8. [8]

    Deep residual learning for image recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. URL https://arxiv.org/abs/1512.03385

    arXiv 2015

Show all 25 references
  1. [9]

    Long short-term memory

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation , 9:1735–80, 12 1997. doi:10.1162/neco.1997.9.8.1735

    1997 doi

  2. [10]

    Empirical evaluation of gated recurrent neural networks on sequence modeling, 2014

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling, 2014. URL https://arxiv.org/abs/1412.3555

    2014 arXiv

  3. [11]

    Transformer model applications: A comprehensive survey and analysis

    Dolly Mittal, Ashish Pant, Palika Jajoo, and Veena Yadav. Transformer model applications: A comprehensive survey and analysis. In Vishal Goar, Aditi Sharma, Jungpil Shin, and M. Firoz Mridha, editors, Deep Learning and Visual Artificial Intelligence, pages 25–38, Singapore, 20...

    2024

  4. [12]

    Hawkins and R

    J. Hawkins and R. Dawkins. A Thousand Brains: A New Theory of Intelligence. Basic Books, 2022. ISBN 9781541675797. URL https://books.google.nl/books?id=U46vzgEACAAJ

    2022

  5. [13]

    Properties of sparse distributed representations and their application to hierarchical temporal memory, 2015

    Subutai Ahmad and Jeff Hawkins. Properties of sparse distributed representations and their application to hierarchical temporal memory, 2015. URL https://arxiv.org/abs/1503.07469

    2015 arXiv

  6. [14]

    The htm spatial pooler—a neocortical algorithm for online sparse distributed coding

    Yuwei Cui, Subutai Ahmad, and Jeff Hawkins. The htm spatial pooler—a neocortical algorithm for online sparse distributed coding. Frontiers in Computational Neuroscience, 11, 2017. ISSN 1662-5188. doi:10.3389/fncom.2017.00111. URL https: //www.frontiersin.org/journals/computati...

    2017

  7. [15]

    A theory of how columns in the neocortex enable learning the structure of the world

    Jeff Hawkins, Subutai Ahmad, and Yuwei Cui. A theory of how columns in the neocortex enable learning the structure of the world. Frontiers in Neural Circuits , 11, 2017. ISSN 1662-5110. doi:10.3389/fncir.2017.00081. URL https: //www.frontiersin.org/journals/neural-circuits/art...

    2017

  8. [16]

    Vicente Reyes-Puerta, Suam Kim, Jyh-Jang Sun, Barbara Imbrosci, Werner Kilb, and Heiko J. Luhmann. High stimulus-related information in barrel cortex inhibitory interneurons. PLOS Computational Biology, 11(6):1–32, 06 2015. doi:10.1371/journal.pcbi.1004121. URL https://doi.org...

    2015 doi

  9. [17]

    Neuronal correlates of local, lateral, and translaminar inhibition with reference to cortical columns

    Moritz Helmstaedter, Bert Sakmann, and Dirk Feldmeyer. Neuronal correlates of local, lateral, and translaminar inhibition with reference to cortical columns. Cerebral Cortex, 19(4):926–937, 10 2008. ISSN 1047-3211. doi:10.1093/cercor/bhn141. URL https://doi.org/10.1093/cercor/bhn141

    2008 doi

  10. [18]

    Neocognitron: A hierarchical neural network capable of visual pattern recognition.Neural Networks, 1(2): 119–130, 1988

    Kunihiko Fukushima. Neocognitron: A hierarchical neural network capable of visual pattern recognition.Neural Networks, 1(2): 119–130, 1988. ISSN 0893-6080. doi:https://doi.org/10.1016/0893-6080(88)90014-7. URL https://www.sciencedirect. com/science/article/pii/0893608088900147

    1988

  11. [19]

    Kwta networks and their applications

    Taek Kwon and Michalis Zervakis. Kwta networks and their applications. Multidimensional Systems and Signal Processing, 6: 333–346, 01 1995. doi:10.1007/BF00983559

    1995 doi

  12. [20]

    Pytorch: An imperative style, high-performance deep learning library, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fan...

    2019 arXiv

  13. [21]

    Harris, K

    Charles R. Harris, K. Jarrod Millman, Stéfan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fe...

    2020

  14. [22]

    Multi30k: Multilingual english-german image descriptions,

    Desmond Elliott, Stella Frank, Khalil Sima’an, and Lucia Specia. Multi30k: Multilingual english-german image descriptions,

  15. [23]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. URL https://arxiv.org/abs/ 1412.6980

    2017 arXiv

  16. [24]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL ’02, page 311–318, USA, 2002. Association for Computational ...

    2002

  17. [2016]

    URL https://arxiv.org/abs/1605.00459

    arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.