arXiv:2602.06797v3 Announce Type: replace-cross Abstract: We study optimal learning rate (LR) schedules under the functional scaling law (FSL) framework (Li et al., 2025), which decomposes training d...
The paper investigates how momentum methods affect large‑batch training in a one‑pass setting using power‑law kernel regression. It derives critical learning rates for SGD, Polyak, and Nesterov, and shows how these rates depend on batch size, momentum, and model capacity. The authors provide scaling laws for risk dynamics, a three‑regime batch‑size phase diagram, and demonstrate that Polyak increases the critical batch size while Nesterov improves data efficiency in the large‑batch regime.
The paper investigates how momentum methods affect large‑batch training in a one‑pass setting using power‑law kernel regression. It derives critical learning rates for SGD, Polyak, and Nesterov, and shows how these rates depend on batch size, momentum, and a capacity exponent. The authors then analyze risk dynamics, optimize final‑step risk under a fixed data budget, and present a three‑regime batch‑size phase diagram that highlights Polyak’s ability to enlarge the critical batch size and Nesterov’s superior data efficiency in the large‑batch regime.
arXiv:2605. 29548v2 Announce Type: replace Abstract: Larger models learn tasks smaller models do not.
arXiv:2606. 29519v1 Announce Type: new Abstract: Long-range learning is hard for recurrent networks trained with stochastic gradient descent, because the influence of a past input fades with the lag $\ell$, and if it fades too fast the dependence cannot be learned from finite data.
arXiv:2605. 24316v3 Announce Type: replace Abstract: Mini-batching is central to large-scale optimization, yet its role in statistical scaling laws remains limited.
arXiv:2512. 22088v3 Announce Type: replace-cross Abstract: The scaling law, a cornerstone of Large Language Model (LLM) development, predicts improvements in model performance with increasing computational resources.
arXiv:2609.36271v1 Announce Type: new Abstract: Recent work has established that power-law spectral conditions on data enable tight convergence bounds for deterministic gradient descent, resolving th...
The paper investigates how much learned memory is required to leverage additional data in autoregressive prediction models. It introduces a predictive‑energy spectrum that jointly governs data and memory scaling, proving a minimax law that links the number of prediction blocks and the size of the learned state to this spectrum. The authors demonstrate that optimal bit allocation and masked query‑key attention mechanisms realize this law, and they provide experimental evidence across multiple pretrained‑model scales.
The paper presents empirical scaling laws for autoregressive language models, linking prediction loss to model size, data size, and compute, and investigates their theoretical basis using a teacher–student linear RNN framework. In this tractable setting, a stable latent linear RNN generates trajectories while a sketched linear recurrent student is trained via full‑batch WSD gradient descent on next‑token prediction. The study derives explicit approximation, optimization, and statistical scaling laws that depend on the sketch dimension, number of trajectories, and trajectory length, revealing how different power‑law exponents for innovation and initialization covariances affect the rates and crossovers between regimes.
The paper presents a spectral theory explaining the phenomenon of grokking, where an initial fit to training data is followed by a delayed improvement in generalization. It shows that for homogeneous networks trained with squared loss and L₂ weight decay, residuals after memorization influence the neural tangent kernel (NTK) dynamics, leading to a transition from lazy to rich learning. The theory predicts that grokking timescales depend on the product of learning rate and weight decay, and that stronger decay can halt fitting, with empirical validation on modular addition tasks using MLPs and Transformers.
Long-range learning is hard for recurrent networks trained with stochastic gradient descent, because the influence of a past input fades with the lag $\ell$, and if it fades too fast the dependence cannot be learned from finite data. This fade is captured by an envelope $f(\ell)$.