Pith. sign in

REVIEW 3 major objections 5 minor 6 cited by

The Surprising Agreement Between Convex Optimization Theory and Learning-Rate Scheduling for Large Model Training

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read A last-iterate bound from non-smooth convex optimization, evaluated at its optimal base learning rate, reproduces the empirical loss curves of cosine and wsd schedules in LLM pretraining and yields transferable learning-rate rules.

desk verdict Solid, honest scheduling paper with real new results; the agreement claim needs one measured quantity (G_t) before it fully lands. read the letter →

arxiv 2501.18965 v2 pith:OUZ7MZ3Y submitted 2025-01-31 cs.LG math.OCstat.ML

classification cs.LGmath.OCstat.ML
keywords learning-ratescheduleswarmup-stable-decaycooldownlast-iterateconvergencenon-smoothconvexoptimizationstochasticgradientdescentlanguagemodelpretrainingtransfer
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 tries to establish that the shapes of learning-rate schedules used in large-model pretraining—cosine and warmup-stable-decay (wsd), including the sudden loss drop when cooldown begins—are reproduced by a worst-case bound from non-smooth stochastic convex optimization. The bound is for the last iterate of SGD, and once its base learning rate is tuned to minimize the bound, the theoretical curve tracks empirical validation-loss curves for transformer training. The paper derives a wsd-specific version of the bound and shows that cooldown removes logarithmic terms, giving a mechanism for the practical benefit of cooldown. It also converts the bound into tuning rules: the optimal base learning rate scales as the inverse square root of the horizon, cosine needs roughly twice the wsd learning rate, and an optimal rate for one cooldown fraction can be transferred to another. Correct, this would mean schedule design and learning-rate transfer for large models can be guided by a first-principles bound rather than by trial and error alone.

What carries the argument

The load-bearing object is the last-iterate schedule bound from Theorem 3.1 (Eq. 6): for iterates $x_{t+1}=x_t-\gamma\eta_t g_t$ with convex losses, it upper-bounds $\mathbb{E}[f(x_T)-f(x^\star)]$ in terms of the initial distance $D$, gradient-norm bounds $G_t$, and the schedule $(\eta_t)$, separated from a base rate $\gamma$. Corollary 3.3 minimizes the bound over $\gamma$, giving $\gamma^\star=\sqrt{T_1/T_2}$ and the plug-in bound $2\sqrt{T_1T_2}$. The paper then evaluates these for cosine and for wsd, which is constant until $T_0$ and then decays linearly; the wsd calculation is what makes cooldown visible as the disappearance of logarithmic terms, and the bound-minimizing rate is what supplies the learning-rate transfer rules.

What would settle it

Train a sufficiently large model, such as a 1B-parameter transformer, with AdamW under cosine and wsd using bound-minimizing base rates, and measure the ratio $\gamma^\star_{\text{cosine}}/\gamma^\star_{\text{wsd}}$ and the size of the cooldown drop; if the ratio departs from roughly 2, the claimed agreement is scale- or optimizer-specific.

Watch

Extended reading notes

Core claim

The central claim is that the last-iterate suboptimality bound for SGD on convex Lipschitz objectives, evaluated at the base learning rate that minimizes it, reproduces the empirical loss curves of cosine and wsd schedules in LLM pretraining. For wsd the paper derives a closed-form bound whose cooldown phase eliminates the logarithmic factor present for constant schedules, explaining the practical cooldown benefit. The same bound predicts that the optimal base learning rate decays like $T^{-1/2}$, that cosine's optimal rate is roughly twice wsd's, and that the ratio of optimal rates across cooldown fractions is stable across horizons; these predictions match re-analysis of real training runs and yield transfer factors such as $\gamma^\star(1)\approx e^{0.7}\gamma^\star(0.2)$ for linear cooldown. The paper also shows the drop during cooldown appears in upper bounds, worst-case lower bounds computed by semidefinite programming, and a two-dimensional non-smooth convex problem, supporting the claim that the phenomenon is not architecture-specific.

Load-bearing premise

The load-bearing premise, flagged by the paper itself in Section 6, is that the schedule-shape behavior of AdamW on non-convex transformer training is governed by a worst-case bound proven for SGD on convex Lipschitz objectives; if that transfer fails at larger scales or for AdamW-specific dynamics, the central agreement and the derived tuning rules lose their foundation.

Editorial extensions

If this is right

  • If the bound is the right testbed, a fully tuned base learning rate makes linear decay the optimal schedule among the studied classes, so the optimal cooldown fraction is one.
  • Continued training can be scheduled by theory: after extending a wsd run from $T_1$ to $T_2$, decreasing the schedule by a computed factor such as $\rho=0.525$ for $T_2=2T_1$ keeps the bound close to a freshly tuned linear-decay run.
  • Learning-rate transfer across schedules becomes a calculable multiplier: with the optimal rate for 20% linear cooldown known, the linear-decay rate is $\gamma^\star(1)\approx e^{0.7}\gamma^\star(0.2)$, avoiding a new sweep.
  • The same logic explains why cosine's cycle length of one is optimal, matching the empirical recommendation for language-model pretraining.
  • The improvement from adapted continued training is worth roughly 6-8% more tokens by the paper's scaling-law estimate, corresponding to about 0.01 validation loss for the 124M and 210M models.

Reading between the lines

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

  • Editorial inference: if the bound remains predictive beyond 210M parameters, schedule experiments could be cheaply screened against the theoretical curve before spending GPU hours, with the bound serving as a prior for candidate schedules.
  • Editorial inference: the paper's analysis suggests a testable decomposition—the cooldown drop is tied to non-vanishing gradient norms; monitoring $\mathbb{E}\|g_t\|^2$ during pretraining could predict whether wsd's drop will be sharp from that curve alone.
  • Editorial inference: the mirror-descent extension in the appendix hints that sign-descent-like preconditioning may admit the same last-iterate bound, which would connect the SGD-based theory to Adam's success without invoking convexity of the full network.
  • Editorial inference: because the bound's schedule-shape predictions are invariant to the scale of $G$ and $D$, a direct test is to vary batch size or loss scaling and check whether the optimal rate ratios remain unchanged.
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 / 5 minor

Summary. The paper studies the connection between a last-iterate suboptimality bound for non-smooth stochastic convex optimization (Defazio et al., 2023) and learning-rate schedules for large model training. It derives a bound for the wsd schedule, shows that the bound's shape over time resembles empirical validation-loss curves for cosine and wsd schedules, and uses the bound's minimizer to construct schedule-extension rules and to transfer optimal base learning-rates between schedules. The experiments include predictions for 124M and 210M Llama-type models, ImageNet ResNet50 with SGD, and OpenWebText2 language models. The paper also provides a mirror-descent extension in Appendix F and a PEP-based lower-bound analysis in Appendix B.4.

Significance. The paper offers a novel perspective on learning-rate scheduling by showing that predictions from a simple non-smooth convex bound can match empirical observations. The transfer rules and schedule-extension scheme are validated out-of-sample on real training runs, and the code is available, which is a genuine strength. However, the central link—from a worst-case SGD bound with an unmeasured gradient-norm profile to AdamW on non-convex transformers—is not fully established. Because the claimed 'surprising agreement' is the main thesis, this gap tempers the significance of the theoretical framing, even though the practical transfer rule (exp(0.7) factor) is a useful empirical contribution.

major comments (3)
  1. [Section 4.2, Fig. 6] The conclusion in Sections 3.1 and 4.1 that the wsd cooldown drop and the cosine/wsd comparison are reproduced by the bound relies on the assumption G_t = 1 for all t. Figure 6 shows that the sudden drop disappears when G_t is proportional to t^alpha for alpha = -0.5 or -1, and the schedule comparison changes materially. Since no gradient-norm measurements are reported for any of the LLM, ImageNet, or OpenWebText2 runs, and since AdamW is invariant to the scale of the stochastic gradient, it remains unclear whether the empirical runs satisfy the required profile. This is a load-bearing assumption for the claimed agreement, and it should be tested by measuring G_t during training or replaced by a more robust theoretical condition.
  2. [Section 6] The manuscript explicitly acknowledges that the theoretical results apply to SGD on convex objectives while all main experiments (Figs. 1, 10, 12 and the OpenWebText2/ImageNet replications) use AdamW or SGD on non-convex models. This gap is more than a caveat because the practical quantities used in Section 5, such as the schedule-reduction factor rho = 0.525 for T2 = 2T1 and the transfer factor exp(0.7) from Fig. 11, are derived from the convex bound. The evidence offered for transfer (the mirror-descent extension in Appendix F and references to SGD/Adam equivalence) is circumstantial. The paper should provide a direct empirical test of whether the bound's schedule-shape predictions hold for AdamW on a small transformer, or present the tuning rules as heuristics rather than theory-based rules.
  3. [Section 3.1, Figs. 1–3] The agreement between the theoretical bound and empirical loss curves is assessed visually, without a quantitative measure. The bound has an arbitrary vertical scale (D = G = 1) and is a worst-case upper bound rather than a model of the loss trajectory. The authors should report a quantitative summary of the match, such as the correlation between the predicted and observed curves after optimally scaling the bound, or a defined feature (e.g., the drop height) with an error bar. Without such a measure, the 'surprisingly close match' claim is difficult to evaluate and could be confounded by the many degrees of freedom in the schedules and the chosen base learning-rates.
minor comments (5)
  1. [Fig. 6 caption] The word 'converegnce' in the caption is a typo and should be corrected to 'convergence'.
  2. [Section 4.2] The phrase 'the bound of the expected gradient norms G1:T' should be 'the bound on the expected gradient norms G1:T'.
  3. [Theorem 3.4] The symbol '≾' is used without a definition; please specify that it denotes an asymptotic inequality as T tends to infinity.
  4. [Section 5.1] The statement 'we verified that changing the values of G, D, or T1 do not affect the result' is not supported by a figure or table; a one-line explanation of the multiplicative scaling would be clearer.
  5. [Appendix B.5] The description of the grid for base learning-rate gamma and cooldown fraction c does not mention the number of seeds or run-to-run variance, so the fitted optimum gamma*(c) in Fig. 12a has no uncertainty estimate.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the bound-based predictions are computed from an externally cited theorem and self-contained algebra, and the empirical comparisons use independent training data; the main limitations are assumption mismatches rather than circular reductions.

full rationale

The paper's derivation chain is not circular on the definition used in this review. Theorem 3.1 is taken from Defazio et al. (2023) and proved in Appendix E, while Corollary 3.3 and the wsd bound (Theorem 3.4/G.1) are new algebra that follows from the stated convexity and gradient-norm assumptions. The schedule-shape predictions (cosine vs. wsd curves, the approximate 2-to-1 optimal learning-rate ratio, the inverse-square-root scaling with horizon, the cooldown log-term cancellation, and the e^{0.7} transfer factor) are computed by minimizing Eq. (8) with G_t = D = 1; because these constants factor out of the minimizer for constant G_t, the predictions do not depend on fitted values. The empirical checks use loss curves from Hägele et al. (2024) and new runs on Llama-style models, ImageNet, and OpenWebText2, and none of these runs is used to set a free parameter in the bound. The clearest limitations, namely SGD theory versus AdamW practice (Section 6) and the unmeasured gradient-norm profile whose decay removes the cooldown drop (Section 4.2), are assumption mismatches rather than reductions of the prediction to its inputs. One mild methodological caveat is that the empirical transfer check in Section 5.2 estimates the optimal learning rate by fitting h_c(γ) = A_c/γ + B_cγ + C_c, whose functional form is borrowed from the bound (8); this makes the estimator share the theory's parametric shape, but the fitted coefficients A_c and B_c are free and data-determined, so the agreement is not forced by construction. Overall circularity is minimal.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The theoretical predictions rest on the standard convex Lipschitz SGD assumptions (A1)-(A3); the transfer to practice additionally assumes that AdamW training of transformers has the same schedule-dependent suboptimality shape as the SGD bound, and that gradient norms do not vanish so that G_t = G is a valid idealization. D and G are set to 1 in simulations but cancel in all transfer predictions. The fitted h_c coefficients are used only to interpolate empirical optimal LRs, not to generate theory.

free parameters (3)
  • Initial distance D = 1
    Set to 1 in all bound simulations; the optimal gamma and all transfer predictions are invariant to D when G is constant, so this is a normalization rather than a fitted parameter.
  • Gradient norm bound G = 1
    Set to 1 and assumed constant over t (G_t = G). This constancy is the modeling choice that produces the cooldown drop; the transfer ratios are invariant to the value of G.
  • h_c coefficients (A_c, B_c, C_c) = fitted per cooldown fraction c
    Used in Section 5.2 to interpolate the empirical optimal learning rate from a coarse grid; the functional form is inspired by the bound (8). These are validation-side fits and do not enter the theoretical predictions.
assumptions (4)
  • domain assumption Each loss f(.,s) is convex (A1)
    Section 3, Eq. (4); the entire bound is proved for convex objectives, while LLM training is non-convex; the paper assumes the schedule-shape behavior carries over without proof (Section 6).
  • domain assumption Expected squared gradient norms are bounded by G_t^2 (A3), with G_t = G constant
    Section 3 and Section 4.2; the cooldown drop in the bound appears only for non-vanishing gradient norms; the constancy is justified by the presence of optimization noise in practice.
  • domain assumption AdamW training follows the same schedule-dependent suboptimality shape as the SGD bound
    Section 6 (most glaring limitation); the empirical loss curves are AdamW, the theory is SGD, and the equivalence is a working hypothesis with only indirect support.
  • domain assumption The convex optimization landscape is a faithful proxy for the transformer loss landscape
    Section 6; the paper states it has no explanation for why convex theory matches, citing work on near-convexity of NN landscapes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Surprising Agreement Between Convex Optimization Theory and Learning-Rate Scheduling for Large Model Training." pith.science (2026). https://pith.science/paper/OUZ7MZ3Y

@misc{pith2026250118965,
  author       = {Pith},
  title        = {Pith review of: The Surprising Agreement Between Convex Optimization Theory and Learning-Rate Scheduling for Large Model Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OUZ7MZ3Y}},
  note         = {Machine review of arXiv:2501.18965}
}
read the original abstract

We show that learning-rate schedules for large model training behave surprisingly similar to a performance bound from non-smooth convex optimization theory. We provide a bound for the constant schedule with linear cooldown; in particular, the practical benefit of cooldown is reflected in the bound due to the absence of logarithmic terms. Further, we show that this surprisingly close match between optimization theory and practice can be exploited for learning-rate tuning: we achieve noticeable improvements for training 124M and 210M Llama-type models by (i) extending the schedule for continued training with optimal learning-rate, and (ii) transferring the optimal learning-rate across schedules.

Figures

Figures reproduced from arXiv: 2501.18965 by the authors.

Figure 1
Figure 1. Strikingly similar: Validation loss for a 210M Llama model trained with AdamW (left) and the theoretical suboptimality bound (6) from convex optimization (right). Both plots show wsd and cosine schedule with different training lengths T, and with base learning-rate of cosine being twice as large as for wsd. by an arguably much simpler schedule, that combines a con￾stant part with a cooldown period in the end (Hu et … view at source ↗
Figure 2
Figure 2. Schedule (left) and theoretical bound (right) for cosine and wsd, and various T, with base learning-rate γ ⋆ . 10−2 10−1 Base learning-rate γ 10−1 100 Final bound Ω T wsd T ∈ [200, 300, 400, 500, 600] cosine T ∈ [200, 300, 400, 500, 600] (a) Learning-rate sweep 200 400 600 Training horizon T 0.02 0.03 0.04 0.05 0.06 0.07 Optimal base learning-rate γ ? cosine, γ ? = 0.92 · T −0.492 wsd, γ ? = 0.47 · T −0.488 (b) Opti… view at source ↗
Figure 3
Figure 3. Optimal base learning-rate decays with inverse square￾root of training horizon T (right). For cosine, it is roughly twice as large as for wsd (as 0.92/0.47 ≈ 2). The dashed curve in the right-hand side plot is obtained with a least-squares fit. this with an ablation study that uses a standard bound on the minimum suboptimality instead; there, the theoretical bound does not resemble empirical loss curves (cf. Appendi… view at source ↗
Figures from the paper (24 more)
Figure 4
Figure 4. Figure 4: (Left) Optimal base learning-rate increases with cooldown fraction. (Right) For fixed γ, the optimal cooldown fraction can be smaller than 1. The analogous curves for real ex￾periments with similar parabola shapes are in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Schedule (left) and theoretical convergence (right) for varying cooldown fraction. With optimal base learning-rate γ ⋆ , starting the cooldown at T0 = 1 is optimal [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 8
Figure 8. Figure 8: (Left) Transferring the wsd schedule from horizon T1 = 4000 to T2 ∈ [1.5T1, 4T1]. (Right) Not adapting the cooldown length leads to significant suboptimality. Dashed horizontal lines mark bound for the linear-decay schedule with tuned γ ⋆ . (B2) Alternatively, we can k…
Figure 7
Figure 7. Figure 7: PEP lower bound matches the upper bound Ωt in shape. 0 5000 10000 15000 Iteration t 0.000 0.005 0.010 0.015 0.020 γ ?ηt 0 5000 10000 15000 Iteration t 0.02 0.04 0.06 0.08 0.10 Ωt linear-decay same base lr+cooldown adapted cooldown [PITH_FULL_IMAGE:figures/full_fig_p00…
Figure 9
Figure 9. Figure 9: Transfering the learning-rate schedule from horizon T1 = 4000 to T2 ∈ [1.5T1, 4T1] (see also [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: Transferring the learning-rate schedule from horizon T1 = 50 000 to T2 ∈ [2T1, 4T1]. Decreasing the base learning￾rate (green) after 40k steps leads to small improvements in vali￾dation loss compared to keeping it the same (grey). We discuss the significance of the di…
Figure 11
Figure 11. Figure 11: Transferring the optimal base learning-rate from cooldown fraction c to linear-decay (c = 1): for linear cooldown (left) and 1-sqrt cooldown (right). Dashed lines are fitted poly￾nomial of degree 6. 0.0 0.2 0.4 0.6 0.8 1.0 Cooldown fraction c 0.0 0.5 1.0 1.5 ln( γ0.99…
Figure 12
Figure 12. Figure 12: (Left) Re-analysis of learning-rate transfer ( [PITH_FULL_IMAGE:figures/full_fig_p008_12.png]
Figure 13
Figure 13. Figure 13: Same as [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]
Figure 14
Figure 14. Figure 14: Same as [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]
Figure 15
Figure 15. Figure 15: Same as [PITH_FULL_IMAGE:figures/full_fig_p013_15.png]
Figure 16
Figure 16. Figure 16: Same as [PITH_FULL_IMAGE:figures/full_fig_p013_16.png]
Figure 17
Figure 17. Figure 17: (Left) Sudden drop of the loss for wsd schedule for a convex, non-smooth problem. (Right) Iterate path for the three schedules. For wsd, the cooldown period is indicated with the dashed line. Star marks solution. B.2. Schedule Comparison We compare the upper bound Ωt …
Figure 18
Figure 18. Figure 18: Comparison of various learning-rate schedules. Convergence is plotted with the optimal base learning-rate γ ⋆ (chosen individually for each schedule). To the best of our knowledge, this recommendation is based mostly on empirical insights. Using the bound obtained in …
Figure 19
Figure 19. Figure 19: Comparison of cycle lengths for the cosine schedule. Compare to Figure A1 in Hoffmann et al. (2022). B.4. Details on Lower Bound Computation We provide additional details for the simulation in Section 4.3. We compute the lower bounds with the PEPit package (Goujaud et…
Figure 20
Figure 20. Figure 20: (Left) The benefit of cooldown is reflected in the absence of logarithmic terms. Dark grey marks the bound of the constant schedule. (Right) Plotting the individual terms of the bound Ωt = T1/γ + γT2 with γ = γ ⋆ for the wsd schedule. The sudden drop of the bound come…
Figure 21
Figure 21. Figure 21: (Left) Analogous to [PITH_FULL_IMAGE:figures/full_fig_p018_21.png]
Figure 22
Figure 22. Figure 22: (Left) Decreasing factor computed with Corollary 3.3 for extended schedule up to T (where T1 is length of the short run). See Section 5.1, (B2) for details. (Right): Extended schedule for the training runs in [PITH_FULL_IMAGE:figures/full_fig_p018_22.png]
Figure 23
Figure 23. Figure 23: (Left) Transferring the 1/sqrt schedule with linear cooldown from horizon T1 = 4000 to T2 ∈ [1.5T1, 4T1]. (Right) Adapting the cooldown length has only small benefits. Dashed horizontal lines mark bound for linear-decay schedule with tuned γ ⋆ . See Section 5.1, (B1) …
Figure 24
Figure 24. Figure 24: Experiment with adapted cooldown length for 124M model (left) and 210M (right). See Section 5.1, (B1) for details. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_24.png]
Figure 25
Figure 25. Figure 25: The theoretical bound (left) for a range of base learning rates γ (and setting D = G = 1) qualitatively matches the empirical (validation) loss curves for training ResNet50 on Imagenet with SGD (middle). (We display a running average over five epoch in thick to smooth…
Figure 26
Figure 26. Figure 26: Training ResNet50 on Imagenet with SGD: Plotting the three best base learning rates γ for both cosine and wsd schedule, as well as linear-decay schedule with learning rate transfer following Section 5.2. Sudden drop of training loss (middle) and increase of validation…
Figure 27
Figure 27. Figure 27: Training three different model sizes on OpenWebText2. We observe the same characteristic drop of the loss for wsd, as well as matching performance of wsd and cosine. In each run the base learning rate of cosine is twice as large as for wsd. C.3. Computing the Bound-mi…
Figure 28
Figure 28. Figure 28: Optimizing the bound ΩT with respect to the schedule η1, . . . , ηT . (Left) Convergence of the value of the bound, matching the bound of a linear-decay schedule. (Right) Optimization trajectory from constant schedule (yellow) to linear-decay (purple). 21 [PITH_FULL_…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

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

  1. A Defense of the Quadratic Model

    cs.LG 2026-07 conditional novelty 7.0 of 10

    Local Taylor-expanded quadratic models reproduce a 150M-parameter LLM's validation loss for up to 10% of training late in the run, and LLM pretraining operates within a factor of 2 of a stochastic or deterministic edg...

  2. Scaling Collapse Reveals Universal Dynamics in Compute-Optimally Trained Neural Networks

    cs.LG 2025-07 conditional novelty 7.0 of 10

    Compute-optimally trained networks of different sizes show loss curves that collapse onto one universal curve after normalization; with learning rate decay, the collapse is tighter than seed-to-seed noise, providing a...

  3. AGNFormer I: Reconstruction of AGN spectra using a probabilistic transformer model

    astro-ph.GA 2026-07 conditional novelty 6.0 of 10

    An uncertainty-aware transformer reconstructs masked AGN broad lines and spectral halves with 4-16% flux errors and beats eleven purpose-built Lyα-reconstruction algorithms on a blind benchmark.

  4. WSqD: A Horizon-Free Learning Rate Schedule for Large Model Training

    cs.LG 2026-07 conditional novelty 6.0 of 10

    WSqD uses a horizon-independent shifted inverse-square-root base plus linear cooldown to get optimal last-iterate rates in convex stochastic optimization and match or beat tuned WSD on LLM pretraining with one reused ...

  5. Enhancing Optimizer Stability: Momentum Adaptation of The NGN Step-size

    cs.LG 2025-08 conditional novelty 6.0 of 10

    NGN-M, a momentum variant of the NGN step-size, provably converges at O(1/sqrt(K)) under milder assumptions and shows wider step-size stability than Adam, Momo, and SGDM in vision and language tasks.

  6. Why Do We Need Warm-up? A Theoretical Perspective

    cs.LG 2025-10 conditional novelty 5.0 of 10

    Under the proposed (H0,H1)-smoothness condition, gradient descent with a warm-up-style adaptive step-size provably converges faster than with any fixed step-size.

Reference graph

Works this paper leans on

30 extracted references · 19 canonical work pages · cited by 6 Pith papers

  1. [1]

    Why you don't overfit, and don't need Bayes if you only train for one epoch

    Aitchison, L. Why you don’t overfit, and don’t need Bayes if you only train for one epoch. arXiv:2411.14478, Novem- ber

  2. [2]

    For validation set metrics, we display a running average over five epoch in thick to smoothen the plot, and the original data in thin

    improves over wsd, which confirms the findings of Section 5.2. For validation set metrics, we display a running average over five epoch in thick to smoothen the plot, and the original data in thin. Note that the final train loss of wsd appears slightly higher as we display the epoch-wise average of mini-batch losses; due to the steep descent of the loss a...

  3. [7]

    Loss landscape characterization of neural networks without over-parametrization

    9 Learning Rate Scheduling for Large Model Training Islamov, R., Ajroldi, N., Orvieto, A., and Lucchi, A. Loss landscape characterization of neural networks without over-parametrization. arXiv:2410.12455, October

  4. [12]

    Rethinking conventional wisdom in machine learn- ing: From generalization to scaling

    Xiao, L. Rethinking conventional wisdom in machine learn- ing: From generalization to scaling. arXiv:2409.15156, September

  5. [13]

    No more Adam: Learning rate scaling at initialization is all you need

    Xu, M., Xiang, L., Cai, X., and Wen, H. No more Adam: Learning rate scaling at initialization is all you need. arXiv:2412.11768, December

  6. [14]

    and Glineur, F

    Zamani, M. and Glineur, F. Exact convergence rate of the last iterate in subgradient methods. arXiv:2307.11134, July

  7. [15]

    • Appendix B: supplementary information on our experiments

    11 Learning Rate Scheduling for Large Model Training Appendix The supplementary material is organized as follows: • Appendix A: ablation on bounds on the minimal suboptimality bounds. • Appendix B: supplementary information on our experiments. • Appendix C: additional experiments for training language and vision models • Appendix D: technical lemmas • App...

  8. [16]

    2, but with Ωt from (11) The bound on the best-so-far bound has a very different shape of the last-iterate bound

    Same as Fig. 2, but with Ωt from (11) The bound on the best-so-far bound has a very different shape of the last-iterate bound. This shows that standard bounds such as in Theorem A.1 do not capture the real-world convergence observed in Hägele et al. (2024). 12 Learning Rate Scheduling for Large Model Training 10□2 10□1 Base learning-rate γ 10□1 100 Final ...

Show all 30 references
  1. [18]

    Convergence is plotted with the optimal base learning-rate γ⋆ (chosen individually for each schedule)

    Comparison of various learning-rate schedules. Convergence is plotted with the optimal base learning-rate γ⋆ (chosen individually for each schedule). To the best of our knowledge, this recommendation is based mostly on empirical insights. Using the bound obtained in Theorem 3....

  2. [19]

    Compare to Figure A1 in Hoffmann et al

    Comparison of cycle lengths for the cosine schedule. Compare to Figure A1 in Hoffmann et al. (2022). B.4. Details on Lower Bound Computation We provide additional details for the simulation in Section 4.3. We compute the lower bounds with the PEPit package (Goujaud et al., 202...

  3. [20]

    Details on Experiments in Fig

    15 Learning Rate Scheduling for Large Model Training B.5. Details on Experiments in Fig. 1a and Section 5 Training details. The loss curves in Fig. 1a are an exact reproduction of the curves in (Hägele et al., 2024, Fig. 3); they are obtained from training a 210M Llama-style t...

  4. [23]

    Dark grey marks the bound of the constant schedule

    (Left) The benefit of cooldown is reflected in the absence of logarithmic terms. Dark grey marks the bound of the constant schedule. (Right) Plotting the individual terms of the bound Ωt = T1/γ + γT2 with γ = γ⋆ for the wsd schedule. The sudden drop of the bound comes from the...

  5. [27]

    We observe the same characteristic drop of the loss for wsd, as well as matching performance of wsd and cosine

    Training three different model sizes on OpenWebText2. We observe the same characteristic drop of the loss for wsd, as well as matching performance of wsd and cosine. In each run the base learning rate of cosine is twice as large as for wsd. C.3. Computing the Bound-minimizing ...

  6. [28]

    Optimizing the bound ΩT with respect to the schedule η1, . . . , ηT . (Left) Convergence of the value of the bound, matching the bound of a linear-decay schedule. (Right) Optimization trajectory from constant schedule (yellow) to linear-decay (purple). 21 Learning Rate Schedul...

  7. [29]

    ≤ Z t 0 1 s + 1ds ≤ tX s=1 1 s = Ht ≤ 1 + ln(t). E. Missing Proofs The following lemma will be the basic inequality for subsequently proving Theorem E.2; it is a standard result in the online learning and convex optimization literature (Zinkevich, 2003). Lemma E.1. Let 1 ≤ k ≤...

  8. [30]

    Theorem 3.1 follows from applying Theorem E.2 with ˆηt := γηt

    If one carries through this term to the end, then we obtain multiple terms in the bound that depend on ηT . Theorem 3.1 follows from applying Theorem E.2 with ˆηt := γηt. We finally prove Theorem E.2. Proof. First, apply Lemma E.1 with u → x⋆ and k → 1 to obtain TX t=1 ηtE[f (...

  9. [31]

    From the (generalized) Cauchy-Schwarz inequality combined with Young’s inequality, we have s3 ≤ µ 2 ∥xt+1 − xt∥2 + η2 t 2µ ∥gt∥2 ∗

    From (17), we have s2 = Bψ(u, xt) − Bψ(u, xt+1) − Bψ(xt+1, xt). From the (generalized) Cauchy-Schwarz inequality combined with Young’s inequality, we have s3 ≤ µ 2 ∥xt+1 − xt∥2 + η2 t 2µ ∥gt∥2 ∗. Using that −Bψ(xt+1, xt) ≤ −µ 2 ∥xt+1 − xt∥2, we obtain ηt[f (xt, st) − f (u, st)...

  10. [400]

    For each schedule we sweep the base learning-rate γ and plot the bound Ωt for γ = γ⋆ obtained from Corollary 3.3. B.3. Cosine Cycle Length For the cosine schedule, an important hyperparameter is its cycle length, that is, the amount of training where the schedule first reaches...

  11. [1970]

    Training trajectories, mini-batch losses and the curious role of the learning rate

    Sandler, M., Zhmoginov, A., Vladymyrov, M., and Miller, N. Training trajectories, mini-batch losses and the curious role of the learning rate. arXiv:2301.02312, January

  12. [1983]

    Optimal linear decay learning rate schedules and further refinements

    Defazio, A., Cutkosky, A., Mehta, H., and Mishchenko, K. Optimal linear decay learning rate schedules and further refinements. arXiv:2310.07831, October

  13. [2003]

    Chinchilla scaling: A replication attempt

    Besiroglu, T., Erdil, E., Barnett, M., and You, J. Chinchilla scaling: A replication attempt. arXiv:2404.10102, April

  14. [2013]

    M., Cox, D

    Shen, Y ., Stallone, M., Mishra, M., Zhang, G., Tan, S., Prasad, A., Soria, A. M., Cox, D. D., and Panda, R. Power scheduler: A batch size and token number agnostic learn- ing rate scheduler. arXiv:2408.13359, August

  15. [2016]

    We train all models with SGD with heavy-ball momentum

    on Imagenet. We train all models with SGD with heavy-ball momentum. Training is done using the timm library (Wightman, 2019). All runs are using weight decay of 0.0001, momentum 0.9, batch size 4 × 256, and standard data augmentation techniques.8 Figs. 25 and 26 confirm our pr...

  16. [2018]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778,

  17. [2019]

    For all further details we refer to Hägele et al

    with a weight decay of 0.1. For all further details we refer to Hägele et al. (2024, App. A.1). Note that all training curves show the validation loss computed over a subset of 32 batches, while the final validation loss is computed over approx. 6 000 batches; hence, the final...

  18. [2020]

    Rockafellar, R

    [Online; accessed 08- November-2024]. Rockafellar, R. T. Convex Analysis. Princeton Mathematical Series, No

  19. [2021]

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-V oss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Lit...

  20. [2022]

    More concretely, assume we have trained a model of sizeN1 for D1 tokens

    states that the loss L(N, D) for a model with parameters N after training for D tokens can be estimated with L(N, D) = E + A N α + B Dβ , (12) where E, A, B, α, βare usually fitted from data. More concretely, assume we have trained a model of sizeN1 for D1 tokens. To arrive at...

  21. [2023]

    B., Hendrickx, J

    Taylor, A. B., Hendrickx, J. M., and Glineur, F. Smooth strongly convex interpolation and exact worst-case per- formance of first-order methods. Mathematical Program- ming, 161(1-2):307–345, 2017a. Taylor, A. B., Hendrickx, J. M., and Glineur, F. Exact worst- case performance ...

  22. [2024]

    Accurate, large minibatch SGD: Training ImageNet in 1 hour

    Goyal, P., Dollár, P., Girshick, R., Noordhuis, P., Wesolowski, L., Kyrola, A., Tulloch, A., Jia, Y ., and He, K. Accurate, large minibatch SGD: Training ImageNet in 1 hour. arXiv:1706.02677, June

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.