Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

EXAdam: The Power of Adaptive Cross-Moments

T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read EXAdam modifies Adam with cross-moment debiasing and a gradient-acceleration term, reporting 38.46% faster convergence and 90.83% CIFAR-10 test accuracy, beating Adam's 89.66%.

desk verdict Genuinely new Adam variant with a clearly specified update rule, but the empirical case is internally inconsistent—impossible loss values, miscomputed margins, and a language-model run that contradicts the generalization claim. read the letter →

arxiv 2412.20302 v2 pith:DLDRIOY7 submitted 2024-12-29 cs.LG cs.AImath.OC

classification cs.LGcs.AImath.OC
keywords Adamoptimizerbiascorrectioncross-momentgradientaccelerationadaptivelearningratesCIFAR-10convergencespeeddeepoptimization
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

EXAdam is a proposed drop-in replacement for the Adam optimizer. Its update rule replaces Adam's bias-corrected moments $\hat m$ and $\hat v$ with cross-moment corrected versions $\tilde m$ and $\tilde v$, and adds a gradient-acceleration term $\tilde g$ so that each parameter update mixes the smoothed momentum with the current gradient. The paper's central claim is that these two changes together converge faster and generalize better than Adam, with the headline evidence a CIFAR-10 CNN run where EXAdam reaches the same training accuracy in 38.46% fewer steps and reaches 90.83% test accuracy versus 89.66% for Adam. If true, this matters because EXAdam adds only about 2.5% computational overhead while improving on Adam, AdamW, and RMSProp in the reported experiments, offering practitioners a cheap way to speed up and improve training without changing model code.

What carries the argument

The load-bearing mechanism is the cross-moment debiasing pair $\tilde m$ and $\tilde v$ together with the gradient accelerator $\tilde g$ in Algorithm 1. The pair couples the first and second moment estimates: $\tilde m$ scales the standard first-moment debias by $1 + \frac{v}{v+\epsilon}\beta_2^t$, and $\tilde v$ scales the second-moment debias by $1 + \frac{m^2}{m^2+\epsilon}\beta_1^t$. Because $\beta_1^t$ and $\beta_2^t$ decay, these factors are largest in the first iterations, so the new terms mainly change early training and then fade toward Adam's corrections. The same variance ratio appears in $\tilde g$, which adds the current gradient directly to the momentum term in the update, so steep regions with large $v$ get more weight from the instantaneous gradient while flat regions rely more on momentum.

What would settle it

Re-run the CIFAR-10 CNN experiment across a grid of learning rates from $10^{-3}$ to $10^{-5}$ with several seeds; if Adam, AdamW, or RMSProp reaches or exceeds EXAdam's 90.83% test accuracy or matches its convergence speed under its own best configuration, the claimed general superiority fails.

Watch

Extended reading notes

Core claim

The paper claims that Adam's independent treatment of the first and second moment estimates leaves bias correction poorly matched to the local loss geometry, and that injecting the cross-moment factors $\left(1 + \frac{v}{v+\epsilon}\beta_2^t\right)$ and $\left(1 + \frac{m^2}{m^2+\epsilon}\beta_1^t\right)$ into the debiased estimates, together with a gradient accelerator $\tilde g = \frac{g}{1-\beta_1^t}\left(1 + \frac{v}{v+\epsilon}\beta_2^t\right)$, yields an update $\theta \leftarrow \theta - \alpha \frac{\tilde m + \tilde g}{\sqrt{\tilde v} + \epsilon}$ that is more responsive during early training and more stable later. As $t \to \infty$, the correction factors tend to $1$, so EXAdam asymptotically reduces to Adam. On CIFAR-10 with a CNN, the paper reports 38.46% faster convergence than Adam, gains of 1.96, 2.17, and 1.17 percentage points in training, validation, and test accuracy, and a test accuracy of 90.83% versus 89.66% for Adam; in a MinGPT text-generation run it reports lower training loss than Adam, AdamW, and AdEMAMix with competitive validation loss among the stable optimizers.

Load-bearing premise

The superiority claim rests on comparing every optimizer at learning rate $10^{-4}$ with default hyperparameters and one seed (1234), and on the unproven assumption that the new cross-moment terms preserve Adam's update stability.

Editorial extensions

If this is right

  • EXAdam can replace Adam in existing training code with small elementwise additions, raising CIFAR-10 test accuracy to 90.83% from 89.66% for Adam and 89.71% for AdamW at the same learning rate and schedule.
  • At the same 100-epoch schedule, EXAdam reaches its final training accuracy about 38% faster than Adam, so practitioners can stop training earlier or spend the saved compute elsewhere.
  • The cross-moment corrections fade as $t$ grows, so EXAdam's early-training behavior differs from Adam but its asymptotic updates coincide with Adam's, making the change most consequential at the start of training.
  • On the MinGPT/Shakespeare task, EXAdam achieves lower training loss than Adam, AdamW, and AdEMAMix with competitive validation loss, suggesting the benefit is not specific to convolutional image models.
  • The reported 2.5% computational overhead is small enough that the optimizer can be used as a drop-in replacement without major throughput cost.

Reading between the lines

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

  • The correction factors $\frac{v}{v+\epsilon}\beta_2^t$ and $\frac{m^2}{m^2+\epsilon}\beta_1^t$ are largest at $t=1$ and decay quickly, so a testable prediction is that EXAdam's advantage is concentrated in the first epochs and would shrink when training starts with a learning-rate warm-up.
  • Because $\tilde g$ is added to $\tilde m$ inside the update, EXAdam is not a pure rescaling of Adam's direction; measuring the cosine similarity between EXAdam and Adam update vectors would isolate whether the gain comes from the cross-moment corrections or from the raw gradient injection.
  • The 38.46% convergence gain is measured under a ReduceLROnPlateau scheduler that reacts to validation loss, so a natural extension is to check whether the gain survives under a fixed or cosine learning-rate schedule, since the scheduler itself can amplify or mask optimizer differences.
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

5 major / 5 minor

Summary. The paper proposes EXAdam, an Adam variant that modifies the debiasing terms (Eq. 1) and adds a gradient-based acceleration term (Eq. 2) to the update rule (Eq. 3). The stated contributions are improved moment estimation, faster convergence, and higher accuracy relative to Adam, supported by a CNN experiment on CIFAR-10 (38.46% faster convergence; 90.83% vs 89.66% test accuracy) and a MinGPT text-generation benchmark against several baseline optimizers. The manuscript also claims to provide a theoretical analysis of the new components. The central issue is that the theory is absent and the empirical evidence is reported with internal inconsistencies and insufficient rigor to support the claims.

Significance. If the claims were true, EXAdam would be a valuable low-overhead drop-in optimizer: the algorithm is simple to implement, the update cost is claimed to be only 2.5% above Adam, and the paper includes pseudocode and comparisons against seven baselines. However, the theoretical analysis promised in the abstract does not exist, the CIFAR-10 loss values are numerically inconsistent with the reported accuracies, and the paper's own MinGPT table contradicts the generalization claim. The empirical advantage over Adam is therefore not established, and the significance of the contribution cannot be assessed from the current evidence.

major comments (5)
  1. [Section 2, Eqs. (1)–(3)] The abstract and introduction promise a 'theoretical analysis,' but Section 2 contains no theorem, lemma, proposition, or proof. Statements such as 'the new terms asymptotically converge to the original Adam bias correction terms' are simple algebra, while claims of 'more stable and robust optimization process,' 'improved initial convergence and stability,' and 'enhanced ability to escape saddle points' are asserted without mathematical support. The convergence and stability properties of the update rule (Eq. 3) are central to the paper's value and need to be stated formally and proved (e.g., under standard smoothness or convexity assumptions, or at least with a bounded-update argument).
  2. [Section 3.1, Table 1] The reported loss values are not compatible with the reported accuracies under a standard cross-entropy loss. For Adam at epoch 1, training accuracy is 39.85%, so at least 60.15% of examples are misclassified; each misclassified example has loss at least ln 2 ≈ 0.693 (its true-class probability cannot exceed 1/2 if it is not the argmax), which already gives an average loss of at least about 0.42. The table reports a training loss of 0.01271. This is off by more than an order of magnitude and suggests the loss curves are rescaled or computed with an unspecified formula. Because the 38.46% faster-convergence claim is read off those loss curves (Figure 1 and Table 1), the central empirical claim cannot be checked without a full clarification and the raw cross-entropy values.
  3. [Section 3.1, paragraph after Table 2] The accuracy-gap calculations in the text are incorrect. From Table 2, EXAdam's testing accuracy exceeds Adam's by 1.17 percentage points, AdamW's by 1.12, and RMSProp's by 0.69, not by the stated '1.96, 2.44, and 2.39 percentage points, respectively.' Please correct these numbers and ensure the abstract and all other claims are consistent with Table 2, and distinguish percentage-point differences from relative percentage changes.
  4. [Section 3.2, Table 3] Table 3 contradicts the claim that EXAdam generalizes better than the stable baselines. At iteration 8000, EXAdam has training loss 0.892 and validation loss 1.571, while Adam, AdamW, and AdEMAMix have validation losses of 1.505, 1.494, and 1.502, respectively. The text's statement that the table 'demonstrates the superior performance of EXAdam compared to Adam, AdamW, SGD with Momentum, and AdEMAMix, as evidenced by its lower loss values on both the training and validation sets' is false on the validation side. EXAdam's training-validation gap is 0.679, versus 0.320 for Adam, indicating more overfitting rather than better generalization.
  5. [Section 3.1] The fair-comparison premise is not met. All optimizers use the same learning rate (0.0001) and a single random seed (1234), with no repeated runs or error bars; Section 3 itself concedes that this learning rate 'may not be optimal for each individual algorithm.' The advantage over Adam is a 1.17 percentage-point test-accuracy difference (90.83 vs 89.66), which could be attributable to the choice of seed or learning rate rather than to the algorithm. The claimed general superiority is therefore not established by the reported experiments; several seeds and a hyperparameter check are required.
minor comments (5)
  1. [Algorithm 1, line 9] The parentheses in the gradient-accelerator term are ambiguous: the text writes (1 + v/(v + εβ_t^2)) while Algorithm 1 line 9 can be read as (1 + v/(v+ε)β_t^2). Use the same bracketing in both places.
  2. [Title and Section 2] The title introduces 'cross-moments,' but the paper never defines a cross-moment or explains how Equations (1)–(2) constitute cross-moment estimation; please define the term or consider a more descriptive title.
  3. [Throughout] The writing alternates between first-person singular ('my experiments,' 'my contribution') and first-person plural ('we'), and includes informal passages such as the hiking analogy; a uniform, formal style would make the paper easier to assess.
  4. [Figures 1–3] The figure captions do not specify axes, whether the curves are plotted from raw batches or averaged, or whether any smoothing is applied; without this information the convergence-speed comparison cannot be reproduced from the figures.
  5. [Section 3.1, Table 1] The loss columns are given to five significant figures even though the loss function is never named; please state explicitly that these are cross-entropy values (or correct them if they are not).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EXAdam's empirical claims are tested against external optimizers, with no fitted parameters, self-citation chain, or uniqueness theorem making the conclusion equivalent to its inputs.

full rationale

EXAdam's central empirical claim (38.46% faster convergence and 90.83% test accuracy on CIFAR-10) is evaluated against Adam, AdamW, RMSProp, AdaDelta, and SGD with Momentum under a fixed training protocol. No parameter is fitted to the reported test outcome, and no result from the author's prior work is invoked to force the algorithm's choice. The update rules in Eqs. (1)-(3) define the proposed algorithm; the prose explanations such as 'When the gradients are noisy, indicated by a large second moment estimate v, the correction factor is closer to 1, which means the bias correction is more aggressive. This makes sense' are qualitative interpretations of the definition, not predictions derived from the equations in a way that reduces to the desired conclusion. There is no uniqueness theorem, no self-citation load-bearing step, and no fitted parameter renamed as a prediction. The stated 'theoretical analysis' is largely heuristic and lacks formal convergence proofs, and the empirical comparison rests on a single seed and one learning rate, but these are rigor and reproducibility concerns rather than circularity. The accuracy and loss comparisons are made against external baselines and are not constructed to equal EXAdam's claimed advantage by definition. Therefore the paper shows no significant circularity.

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

No parameters are fitted to data; all optimizers share the same hand-set hyperparameters (alpha = 1e-4, default betas, epsilon = 1e-8, scheduler), which are experimental protocol rather than fitted values. The algorithmic terms m-tilde, v-tilde, and g-tilde are new update components, but they are not physical entities with independent falsifiable predictions, so no invented entities are listed.

assumptions (3)
  • domain assumption Adam's independent debiasing of first and second moments is a limitation that causes suboptimal updates.
    Section 2.1 asserts this without proof or reference; it motivates the new cross-moment terms.
  • ad hoc to paper EXAdam's modified debiasing and added gradient term preserve or improve Adam's convergence and stability.
    No convergence theorem, bound, or formal analysis is given; Section 2.2 and 2.3 provide only verbal claims and a worked numerical example.
  • domain assumption Comparing all optimizers at the same learning rate and without per-optimizer tuning is a fair test of superiority.
    Section 3 states 'using the same learning rate for all optimizers facilitates comparison but may not be optimal for each individual algorithm'; this limits the external validity of the comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EXAdam: The Power of Adaptive Cross-Moments." pith.science (2026). https://pith.science/paper/DLDRIOY7

@misc{pith2026241220302,
  author       = {Pith},
  title        = {Pith review of: EXAdam: The Power of Adaptive Cross-Moments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DLDRIOY7}},
  note         = {Machine review of arXiv:2412.20302}
}
abstract

This paper introduces EXAdam ($\textbf{EX}$tended $\textbf{Adam}$), a novel optimization algorithm that builds upon the widely-used Adam optimizer. EXAdam incorporates two key enhancements: (1) new debiasing terms for improved moment estimation and (2) a gradient-based acceleration mechanism for increased responsiveness to the current loss landscape. These innovations work synergistically to address limitations of the original Adam algorithm, potentially offering improved convergence properties, enhanced ability to escape saddle points, and potentially greater robustness to hyperparameter choices, though this requires further investigation. We provide a theoretical analysis of EXAdam's components and their interactions, highlighting the algorithm's potential advantages in navigating complex optimization landscapes. Empirical evaluations demonstrate EXAdam's superiority over Adam, achieving 38.46% faster convergence and yielding improvements of 1.96%, 2.17%, and 1.17% in training, validation, and testing accuracies, respectively, when applied to a CNN trained on the CIFAR-10 dataset. While these results are promising, further empirical validation across diverse tasks is essential to fully gauge EXAdam's efficacy. Nevertheless, EXAdam represents a significant advancement in adaptive optimization techniques, with promising implications for a wide range of machine learning applications. This work aims to contribute to the ongoing development of more efficient, adaptive, and universally applicable optimization methods in the field of machine learning and artificial intelligence.

Figures

Figures reproduced from arXiv: 2412.20302 by the authors.

Figure 1
Figure 1. The training performance of EXAdam, Adam, AdamW, SGD with momentum, RMSProp, and AdaDelta [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. The validation performance of EXAdam, Adam, AdamW, SGD with momentum, RMSProp, and AdaDelta [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Training loss of MinGPT using EXAdam, Adam, AdamW, AdaFactor, SGD with Momentum, AdEMAMix, [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

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. OmniOpt: Taxonomy, Geometry, and Benchmarking of Modern Optimizers

    cs.LG 2026-07 conditional novelty 6.0 of 10

    A meta-pipeline plus LMO four-axis view yields a dual taxonomy of 108 optimizers, and a multi-objective LLM/vision benchmark shows no single family dominates the quality–cost–memory frontier.

Reference graph

Works this paper leans on

22 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [1]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. 2017

  2. [2]

    The cifar-10 dataset, 2014

    Alex Krizhevsky, Vinod Nair, Geoffrey Hinton, et al. The cifar-10 dataset, 2014

  3. [3]

    Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht

    Ashia C. Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht. The marginal value of adaptive gradient methods in machine learning, 2018

  4. [4]

    On large-batch training for deep learning: Generalization gap and sharp minima, 2017

    Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima, 2017

  5. [5]

    Sharp minima can generalize for deep nets, 2017

    Laurent Dinh, Razvan Pascanu, Samy Bengio, and Yoshua Bengio. Sharp minima can generalize for deep nets, 2017

  6. [6]

    A. S. Nemirovskii and Yurii Nesterov. Optimal methods of smooth convex minimization. Ussr Computational Mathematics and Mathematical Physics, 25:21–30, 1986

  7. [7]

    Dahl, and Geoffrey E

    Ilya Sutskever, James Martens, George E. Dahl, and Geoffrey E. Hinton. On the importance of initialization and momentum in deep learning. In International Conference on Machine Learning, 2013

  8. [8]

    On the momentum term in gradient descent learning algorithms

    Ning Qian. On the momentum term in gradient descent learning algorithms. Neural networks : the official journal of the International Neural Network Society, 12:145–151, 02 1999

Show all 22 references
  1. [9]

    An overview of gradient descent optimization algorithms

    Sebastian Ruder. An overview of gradient descent optimization algorithms. ArXiv, abs/1609.04747, 2016

  2. [10]

    Why momentum really works

    Gabriel Goh. Why momentum really works. 2017

  3. [11]

    Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude

    Tijmen Tieleman and Geoffrey Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 4:26–31, 2012

  4. [12]

    Reddi, Satyen Kale, and Sanjiv Kumar

    Sashank J. Reddi, Satyen Kale, and Sanjiv Kumar. On the convergence of adam and beyond. arXiv preprint arXiv:1904.09237, 2019

  5. [13]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019

  6. [14]

    Adaptive gradient methods with dynamic bound of learning rate

    Liangchen Luo, Yuanhao Xiong, Yan Liu, and Xu Sun. Adaptive gradient methods with dynamic bound of learning rate. ArXiv, abs/1902.09843, 2019

  7. [15]

    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 K ¨opf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu F...

  8. [16]

    Matthew D. Zeiler. Adadelta: An adaptive learning rate method, 2012

  9. [17]

    Improving language understanding by generative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018

  10. [18]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023

  11. [19]

    Adafactor: Adaptive learning rates with sublinear memory cost, 2018

    Noam Shazeer and Mitchell Stern. Adafactor: Adaptive learning rates with sublinear memory cost, 2018

  12. [20]

    The ademamix optimizer: Better, faster, older, 2024

    Matteo Pagliardini, Pierre Ablin, and David Grangier. The ademamix optimizer: Better, faster, older, 2024

  13. [21]

    signsgd: Compressed optimisation for non-convex problems, 2018

    Jeremy Bernstein, Yu-Xiang Wang, Kamyar Azizzadenesheli, and Anima Anandkumar. signsgd: Compressed optimisation for non-convex problems, 2018. 12 EXAdam: The Power of Adaptive Cross-Moments A PREPRINT

  14. [22]

    Deconstructing what makes a good optimizer for language models, 2024

    Rosie Zhao, Depen Morwani, David Brandfonbrener, Nikhil Vyas, and Sham Kakade. Deconstructing what makes a good optimizer for language models, 2024. 13

Pith tools

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