Pith. sign in

REVIEW 3 major objections 4 minor 33 references

Capacity Matters: a Proof-of-Concept for Transformer Memorization on Real-World Data

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Embedding width drives transformer memorization, not depth.

desk verdict A worthwhile proof-of-concept: the SNOMED memorization pipeline and the 'width beats depth' observation are real, but the parameter-equalization formula is wrong and the per-parameter recommendation overreaches. read the letter →

arxiv 2506.14704 v1 pith:UIZHQD2Y submitted 2025-06-17 cs.CL

Anton Changalidis , Aki Härmä This is my paper · ORCID
classification cs.CL
keywords transformermemorizationembeddingsizemodeldepthactivationfunctionsknowledgegraphdataMaximumAttainableCapacitymedicalontologystructuredsequences
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 argues that, for small generative transformers trained to memorize structured medical knowledge, the size of the embedding layer is the main determinant of how fast and how much the model learns. Adding transformer layers without widening the embedding neither speeds up convergence nor improves final memorization, and on simple datasets extra depth can even hurt. A softer activation function, Softmax, is reported to give more stable training and higher memorization capacity than ReLU-style activations. The result matters because it suggests practical design rules for small on-device models that must reliably store domain facts, such as medical assistants on wearables.

What carries the argument

Maximum Attainable Capacity (MAC), a metric that counts the maximum number of training samples a model can memorize when given a large library. It is used to compare architectures by training on datasets derived from a medical ontology, either as static triplets or as longer node-edge sequences, and reporting the number of correctly predicted concepts. The comparison across depths is carried out by a parameter-equalization formula, embedding_size = floor(base_number_of_parameters / n_layers), intended to keep total parameters constant while varying depth and width.

What would settle it

Count the total parameters, including the vocabulary embedding matrix, for a one-layer and a two-layer configuration that the paper's formula treats as equal in size. If the totals differ and memorization capacity tracks the total rather than the embedding size, the claim that embedding size is the key factor is not established.

Watch

Extended reading notes

Core claim

Using synthetic text datasets built from a medical knowledge graph, the authors trained hundreds of small decoder-only transformers and measured memorization with the Maximum Attainable Capacity (MAC) method. Their central claim is that embedding size is the key factor influencing learning speed and final capacity, while increasing the number of layers without increasing embedding size neither accelerates convergence nor improves capacity. They also report that Softmax activation is more stable and achieves higher capacity than ReLU, GELU, or RReLU, and that models trained on longer graph-traversal sequences reach near-perfect memorization faster than models trained on isolated triplets, despite showing more fluctuation during training.

Load-bearing premise

The comparison assumes that the formula used to shrink the embedding size as layers are added actually keeps total model size equal; if it does not, the observed advantage of wider embeddings could just be a total-parameter effect.

Editorial extensions

If this is right

  • If embedding size is the binding factor, then for a fixed parameter budget, designers of small memorization-focused transformers should prefer shallow models with wide embeddings over deep models with narrow embeddings.
  • Adding layers to a small transformer may slow training and reduce final memorization on simple structured data, so depth should not be assumed beneficial for memory-bound tasks.
  • Softmax activations appear to provide more stable training and higher capacity than ReLU-style activations on this type of structured memorization task, contrary to some earlier findings on other tasks.
  • Longer structured sequences, such as graph traversal paths, can be memorized faster and more completely than isolated triplets, suggesting that data structure itself is a lever for capacity.
  • Larger datasets can lead to higher final capacity but require longer warm-up, so there is a trade-off between dataset size and training time that depends on the target capacity.

Reading between the lines

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

  • My inference: if embedding width is the true limiting factor, then capacity per parameter could be improved further by compressing or sharing the embedding matrix, since the embedding dominates parameter count with a fixed vocabulary.
  • My inference: the finding suggests a testable design rule for edge deployment: hold embedding size fixed, prune or remove layers, and measure memorization; the paper's results predict little or no capacity loss.
  • My inference: the contrast with studies reporting ReLU superiority indicates that activation-function effects are task-dependent; a useful next experiment would vary dataset complexity continuously and trace where the advantage flips.
  • My inference: the reported capacity barrier at large dataset sizes with very small embeddings suggests an interaction between vocabulary coverage and embedding rank, which could be probed by measuring the effective rank of the embedding matrix during training.
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

3 major / 4 minor

Summary. This paper presents a proof-of-concept study of transformer memorization on structured data derived from the SNOMED medical ontology, using synthetic triplet and sequence datasets with controlled tokenization. The authors train small decoder-only transformers and report that embedding size is the primary determinant of learning speed and memorization capacity, that adding layers does not help (and sometimes hurts), that Softmax activation is more stable than ReLU/GELU/RReLU, and that larger datasets improve final memorization. The paper introduces a pipeline for converting ontologies into memorization benchmarks and applies the Maximum Attainable Capacity (MAC) metric, claimed to correlate with Maximum Library Size from prior work. The main conclusion is that for simple structured memorization, wider embeddings are preferable to deeper architectures.

Significance. If the central claim holds, the paper provides actionable guidance for designing small on-device transformers for domain-specific memorization, a relevant issue for medical and edge applications. The strengths are the reproducible code release, the use of a real-world ontology with clear ground-truth memorization tasks, repeated seeds with reported variability, and a systematic sweep over depth, width, and activation functions. The paper is also honest about its scope as a proof-of-concept and lists limitations. However, the quantitative claim that embedding size, not depth or total parameter count, drives capacity rests on a parameter-equalization procedure that does not actually equalize total parameters, so the main architectural conclusion is not yet established. The capacity metric also relies on a self-cited prior validation rather than independent evidence.

major comments (3)
  1. [2.3.1 and 3.3] The parameter-equalization formula does not actually keep total model parameters constant. The paper states embedding_size = floor(base_number_of_parameters / n_layers), with base values 16/32/64/128, and claims this ensures fair comparisons. But with a fixed vocabulary, the embedding matrix contributes vocab_size × d_model parameters, so halving d_model when doubling the layer count roughly halves the embedding parameters, and the additional layer adds only its own attention and feed-forward parameters. The paper itself notes that model sizes are 'primarily varying with embedding size and layer count, but also influenced by vocabulary size.' Consequently, the configurations compared in Section 3.3 differ in total parameter count, and the observed advantage of larger embeddings over deeper models could be driven by total parameter count or raw embedding width rather than an architectural law. This is load-bearing for the central claim that 'embedding size is the key factor influencing learning speed' and that depth does not matter. I request the authors report the actual total parameter counts for each configuration and either redesign the comparison to truly equalize total parameters or restrict the claim to what the current design supports, e.g., 'adding layers while keeping embedding size fixed does not improve memorization.'
  2. [3.3 and Table 3] The 'final capacity' claims, especially the capacity barrier at 100,000 samples for the two-layer model with embedding size 8, are not supported by evidence that all configurations have converged. The training is stopped at a fixed 500 epochs, and no plateau diagnostics (e.g., flat accuracy or loss curves, or longer training runs) are shown for the smallest embedding configurations. The lower capacity of the 16-parameter two-layer model (85,935 vs. ~88,200 for other configurations) may reflect slower convergence rather than a true capacity limit. Since the paper distinguishes 'learning speed' from 'final capacity,' the final-capacity statements need either longer training for the small-embedding configurations or an explicit demonstration of saturation.
  3. [2.2 and 3.2] The capacity metric MAC is justified solely by a self-cited prior paper (Härmä et al., 2024) asserting a strong correlation between MLS and MAC; no independent benchmark or in-paper validation is provided. Given that 'capacity' is central to the title and conclusions, the authors should either provide a direct validation of MAC on at least a few configurations (e.g., compare MAC with an MLS-style measurement on the same data) or clearly state that MAC is used as a proxy for memorization accuracy on a large training set, not as an independently validated capacity bound.
minor comments (4)
  1. [2.3.1] The variable name 'base_number_of_parameters' is misleading, since the formula sets the embedding dimension, not a true parameter count. The text should be rephrased to avoid the impression that total parameters are held constant.
  2. [Figure 4 caption] The caption states that 'number of parameters is a total number for all layers: green –16, blue –32', but these values are the base numbers used in the formula, not the actual total parameter count of the models. This should be corrected to avoid confusion.
  3. [2.2] There is a typo: 'trainded' should be 'trained'.
  4. [3.3] The sentence 'a one-layer model with 16 parameters (embedding size is 16, light green)' is incorrect: the model has far more parameters than 16 because the embedding matrix size is vocab_size × 16. The terminology should distinguish the base value from the actual parameter count.

Circularity Check

1 steps flagged · score 4.0 of 10

Self-cited MAC-MLS correlation is load-bearing, but the central embedding-size conclusion is empirical and not reduced to its inputs.

  1. self citation load bearing [Section 2.2, 'Transformers training' (MAC metric definition and justification)]
    "MAC is a computationally efficient alternative to the Maximum Library Size (MLS) method. While MLS involves iteratively training models on progressively larger datasets to determine the largest library size that can be fully memorized, MAC is measuring the maximum number of samples that a model can memorize, provided with a large library. Previous research has shown a strong correlation between MLS and MAC (Härmä et al., 2024), making MAC an effective and time-efficient choice for this study."

    The validity of MAC as the paper's central capacity measure is justified entirely by a correlation result in Härmä et al. (2024), a preprint co-authored by one of the present authors. That prior result is not reproduced, code-verified, or independently benchmarked in this paper, so all capacity conclusions rest on an internally sourced premise rather than on external evidence. The architectural conclusion that embedding size, not depth, drives learning speed is not itself derived from this premise; it comes from direct training curves. But the metric in which 'capacity' is reported is load-bearing for the paper's framing, making this a genuine self-citation dependency rather than a merely cosmetic reference.

full rationale

I found no step in which a prediction is equivalent to an input by construction. The data pipeline, training runs, and accuracy measurements are empirical: accuracy is directly read from token predictions, and capacity values are simply the number of correct predictions (or accuracy times dataset size). No parameter is fitted to a subset of data and then renamed as a prediction. The main claim that embedding size dominates learning speed is supported by comparing convergence curves across configurations; those curves could have come out differently, so the finding is not forced by the experimental design. The parameter-equalization formula in Section 2.3.1 is questionable as a true total-parameter control for a fixed vocabulary, because the embedding matrix scales with vocab_size times d_model while the formula only fixes d_model times n_layers; however, this is a validity/confound concern, not a circularity, since the observed convergence similarity is an empirical result rather than a definitional equivalence. The one identifiable circularity-adjacent element is the self-cited justification of the MAC metric via Härmä et al. (2024): the correlation with MLS is asserted from the authors' own prior work without independent verification in this manuscript. Because the central architectural results also stand on reported accuracy, the paper retains independent content, but the metric on which the capacity narrative rests is internally justified. This warrants a moderate score of 4 rather than a higher one, and the methodological formula issue should be treated as correctness risk, not as circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper adds no new theoretical entities or fitted constants. Its load-bearing choices are the MAC metric, the fixed epoch budget, and the d_model scaling formula. The MAC metric is adopted from the authors' own prior work, and the formula's validity is the main control issue.

free parameters (2)
  • d_model scaling formula for depth comparisons = d_model = floor(base / n_layers), base in {16, 32, 64, 128}
    Hand-chosen formula intended to keep parameter counts equal across layer counts; it does not account for the vocabulary-sized embedding matrix, so the central depth-versus-width comparison rests on this unverified construction.
  • Fixed training budget = 500 or 1000 epochs depending on setup
    Capacity is reported at the end of a fixed budget; claims about maximum attainable capacity assume this budget reaches a plateau for every configuration, which is not established for the larger datasets.
assumptions (4)
  • domain assumption MAC measured as the number of correct predictions at the final checkpoint equals the maximum attainable memorization capacity.
    Stated in Section 2.2 and relied on throughout; the validation is a correlation study in the self-cited Härmä et al. 2024 preprint.
  • ad hoc to paper The formula d_model = floor(base / n_layers) equalizes total model parameters across compared configurations.
    Invoked in Section 2.3.1 to attribute performance differences to architecture alone; the formula only scales one hyperparameter and is not shown to equalize the full parameter count.
  • domain assumption Randomly sampled triplets and BFS-limited sequences preserve enough of the SNOMED graph's structure to count as realistic structured data.
    Sections 2.1.2 and 2.1.3; random choice of one related concept per pair and unique-edge traversal simplify the graph to deterministic lookup tasks.
  • domain assumption Unspecified PyTorch and HuggingFace Transformers defaults do not materially affect the comparisons.
    Section 2.2 says 'All other were default unless specified'; no seed management or full configuration file is shown in the text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Capacity Matters: a Proof-of-Concept for Transformer Memorization on Real-World Data." pith.science (2026). https://pith.science/paper/UIZHQD2Y

@misc{pith2026250614704,
  author       = {Pith},
  title        = {Pith review of: Capacity Matters: a Proof-of-Concept for Transformer Memorization on Real-World Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UIZHQD2Y}},
  note         = {Machine review of arXiv:2506.14704}
}
read the original abstract

This paper studies how the model architecture and data configurations influence the empirical memorization capacity of generative transformers. The models are trained using synthetic text datasets derived from the Systematized Nomenclature of Medicine (SNOMED) knowledge graph: triplets, representing static connections, and sequences, simulating complex relation patterns. The results show that embedding size is the primary determinant of learning speed and capacity, while additional layers provide limited benefits and may hinder performance on simpler datasets. Activation functions play a crucial role, and Softmax demonstrates greater stability and capacity. Furthermore, increasing the complexity of the data set seems to improve the final memorization. These insights improve our understanding of transformer memory mechanisms and provide a framework for optimizing model design with structured real-world data.

Figures

Figures reproduced from arXiv: 2506.14704 by the authors.

Figure 1
Figure 1. Algorithms of triplets (A) and sequences (B) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Trends in training accuracy (upper) and capacity (lower) for the first setup (different data sizes, for triplets [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Trends in training capacity for the second setup (different data sizes, activation functions, and numbers of [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Trends in training accuracy (upper) and capacity (lower) for the third setup (different data sizes, numbers [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Trends in training capacity for the fourth setup (different data sizes, activation functions, and numbers of [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Trends in training accuracy for the second setup (different data sizes, activation functions, and numbers of [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Trends in training accuracy for the fourth setup (different data sizes, activation functions, and numbers of [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 10 canonical work pages

  1. [1]

    Abien Fred Agarap. 2019. https://arxiv.org/abs/1803.08375 Deep learning using rectified linear units (relu) . Preprint, arXiv:1803.08375

    arXiv 2019

  2. [2]

    Zeyuan Allen-Zhu and Yuanzhi Li. 2024. https://arxiv.org/abs/2309.14316 Physics of language models: Part 3.1, knowledge storage and extraction . Preprint, arXiv:2309.14316

    arXiv 2024

  3. [3]

    Simone Balloccu, Ehud Reiter, Vivek Kumar, Diego Reforgiato Recupero, and Daniele Riboni. 2024. https://doi.org/10.48550/arXiv.2401.08420 Ask the experts: sourcing high-quality datasets for nutritional counselling through Human - AI collaboration . arXiv preprint. ArXiv:2401.08420 [cs]

    work page Pith review arXiv doi:10.48550/arxiv.2401.08420 2024

  4. [4]

    Ludwig Boltzmann. 1868. Studien über das gleichgewicht der lebendigen kraft zwischen bewegten materiellen punkten. Wiener Berichte, 58:517--560. Studies on the balance of living force between moving material points

    work page

  5. [5]

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nicholas L. Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zach Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E. Burke, Tristan Hume, Shan Carter, Tom Henigha...

    work page 2023

  6. [6]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin...

    arXiv 2020

  7. [7]

    Wenlin Chen and Hong Ge. 2024. https://arxiv.org/abs/2305.15912 Neural characteristic activation analysis and geometric parameterization for relu networks . Preprint, arXiv:2305.15912

    work page Pith review arXiv 2024

  8. [8]

    Shaker El-Sappagh, Francesco Franda, Farman Ali, and Kyung-Sup Kwak. 2018. https://doi.org/10.1186/s12911-018-0651-5 Snomed ct standard ontology based on the ontology for general medical science . BMC Medical Informatics and Decision Making, 18(1):76

    doi:10.1186/s12911-018-0651-5 2018

Show all 33 references
  1. [9]

    Fan Feng, Feitong Tang, Yijia Gao, Dongyu Zhu, Tianjun Li, Shuyuan Yang, Yuan Yao, Yuanhao Huang, and Jie Liu. 2022. https://doi.org/10.1093/nar/gkac957 Genomickb: a knowledge graph for the human genome . Nucleic Acids Research, 51(D1):D950--D956

    2022 doi

  2. [10]

    Jingwen Fu, Tao Yang, Yuwang Wang, Yan Lu, and Nanning Zheng. 2024. https://arxiv.org/abs/2309.06054 Breaking through the learning plateaus of in-context learning in transformer . Preprint, arXiv:2309.06054

    2024 arXiv

  3. [11]

    Bhumika Gupta, Pralaypati Ta, Keerthi Ram, and Mohanasankar Sivaprakasam. 2024. https://doi.org/10.1109/CIBCB58642.2024.10702112 Comprehensive Modeling and Question Answering of Cancer Clinical Practice Guidelines using LLMs . In 2024 IEEE Conference on Computational Intellige...

    2024

  4. [12]

    Shwai He, Guoheng Sun, Zheyu Shen, and Ang Li. 2024. https://arxiv.org/abs/2406.15786 What matters in transformers? not all attention is needed . Preprint, arXiv:2406.15786

    2024 arXiv

  5. [13]

    Dan Hendrycks and Kevin Gimpel. 2023. https://arxiv.org/abs/1606.08415 Gaussian error linear units (gelus) . Preprint, arXiv:1606.08415

    2023 arXiv

  6. [14]

    Aki Härmä, Marcin Pietrasik, and Anna Wilbik. 2024. https://arxiv.org/abs/2407.15425 Empirical capacity model for self-attention neural networks . Preprint, arXiv:2407.15425

    2024 arXiv

  7. [15]

    Yue Ju, Alka Isac, and Yimin Nie. 2021. https://arxiv.org/abs/2112.15087 Chunkformer: Learning long time series with multi-stage chunked transformer . Preprint, arXiv:2112.15087

    2021 arXiv

  8. [16]

    Tokio Kajitsuka and Issei Sato. 2024. https://arxiv.org/abs/2409.17677 Optimal memorization capacity of transformers . Preprint, arXiv:2409.17677

    2024 arXiv

  9. [17]

    Jisu Kim and Juhwan Lee. 2024. https://arxiv.org/abs/2405.07490 Strategic data ordering: Enhancing large language model performance through curriculum learning . Preprint, arXiv:2405.07490

    2024 arXiv

  10. [18]

    Junghwan Kim, Michelle Kim, and Barzan Mozafari. 2023. https://openreview.net/forum?id=8JCg5xJCTPR Provable memorization capacity of transformers . In The Eleventh International Conference on Learning Representations

    2023

  11. [19]

    Kingma and Jimmy Ba

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

    2017 arXiv

  12. [20]

    Jean-Baptiste Lamy. 2017. https://doi.org/10.1016/j.artmed.2017.07.002 Owlready: Ontology-oriented programming in python with automatic classification and high level constructs for biomedical ontologies . Artificial Intelligence in Medicine, 80:11--28

    2017 doi

  13. [21]

    Sadegh Mahdavi, Renjie Liao, and Christos Thrampoulidis. 2024. https://arxiv.org/abs/2306.02010 Memorization capacity of multi-head attention in transformers . Preprint, arXiv:2306.02010

    2024 arXiv

  14. [22]

    Inyoung Paik and Jaesik Choi. 2023. https://arxiv.org/abs/2304.11692 The disharmony between bn and relu causes gradient explosion, but is offset by the correlation between activations . Preprint, arXiv:2304.11692

    2023 arXiv

  15. [23]

    Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic differentiation in pytorch

    2017

  16. [24]

    Gonçalo Paulo, Stepan Shabalin, and Nora Belrose. 2025. https://arxiv.org/abs/2501.18823 Transcoders beat sparse autoencoders for interpretability . Preprint, arXiv:2501.18823

    2025 arXiv

  17. [25]

    Ahsan Shehzad, Feng Xia, Shagufta Abid, Ciyuan Peng, Shuo Yu, Dongyu Zhang, and Karin Verspoor. 2024. https://arxiv.org/abs/2407.09777 Graph transformers: A survey . Preprint, arXiv:2407.09777

    2024

  18. [26]

    Kai Shen, Junliang Guo, Xu Tan, Siliang Tang, Rui Wang, and Jiang Bian. 2023. https://arxiv.org/abs/2302.06461 A study on relu and softmax in transformer . Preprint, arXiv:2302.06461

    2023 arXiv

  19. [27]

    Daouda Sow, Herbert Woisetschläger, Saikiran Bulusu, Shiqiang Wang, Hans-Arno Jacobsen, and Yingbin Liang. 2025. https://arxiv.org/abs/2502.06733 Dynamic loss-based sample reweighting for improved large language model pretraining . Preprint, arXiv:2502.06733

    2025 arXiv

  20. [28]

    SURF . 2024. https://servicedesk.surf.nl/wiki/display/WIKI/Snellius Snellius - the dutch national supercomputer

    2024

  21. [29]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \' e mi Louf, Morgan Funtowicz, and Jamie Brew. 2019. https://arxiv.org/abs/1910.03771 Huggingface's transformers: State-of-the-art natural language processin...

    2019 arXiv

  22. [30]

    Jinlin Wu, Xusheng Liang, Xuexue Bai, and Zhen Chen. 2024. https://doi.org/10.48550/arXiv.2412.05187 SurgBox : Agent - Driven Operating Room Sandbox with Surgery Copilot . arXiv preprint. ArXiv:2412.05187 [cs]

    2024 doi arXiv

  23. [31]

    Bing Xu, Naiyan Wang, Tianqi Chen, and Mu Li. 2015. https://arxiv.org/abs/1505.00853 Empirical evaluation of rectified activations in convolutional network . Preprint, arXiv:1505.00853

    2015 arXiv

  24. [32]

    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...

  25. [33]

    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 gl...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.