Publications
In reverse chronological order. Preprints and published work.
2026
- PreprintItay Lavie, Clarissa Lauditi, and Cengiz PehlevanIn submission, 2026
A recurring design principle in modern optimizers is to decouple update magnitude from the raw gradient norm, yet its consequences for learning-curve and resource scaling remain unclear. We isolate this mechanism by studying normalized SGD in a random-feature model with power-law teacher and data covariance. Fixed-norm updates induce an effective learning rate that grows as gradients shrink. We derive a dynamical mean-field theory (DMFT) describing the joint dependence of the loss on training time, model width and batch size. Normalization initially accelerates SGD, mapping the power-law exponent r_SGD<1 to 2r_SGD/(1-r_SGD), with exponential convergence at r_SGD=1 and formal finite-time convergence for r_SGD>1. At finite step size, however, the same feedback ultimately breaks the acceleration and leads to marginal stability. The late-time theory yields width-limited, edge-of-stochastic-stability (EoSS), and deterministic edge-of-stability (EoS) regimes. These phases determine when larger batches or wider models reduce serial training time at comparable compute. We quantify in which of these phases increased batch size or width can compensate the excess compute use per step by fewer optimization steps to target loss. Linearized ResNet experiments on CIFAR-5M support the predicted acceleration, breakdown, and resource-scaling trends. Together, these results connect normalization-induced acceleration, EoS effects, and width–batch allocation within a solvable theory.
@misc{lavie2026solvable, title = {A Solvable Model of Adaptive Learning Rate Rescaling: Acceleration, Stability \& Scaling}, author = {Lavie, Itay and Lauditi, Clarissa and Pehlevan, Cengiz}, year = {2026}, publisher = {arXiv}, } - NeurIPS-WItay Lavie, Clarissa Lauditi, and Cengiz PehlevanIn NeurIPS 2026 Workshop on Optimization for Machine Learning (OPT 2026), 2026
Selected for an oral presentation at OPT 2026 (<5% of accepted papers).
A recurring design principle in modern optimizers is to decouple update magnitude from the raw gradient norm, ranging from coordinate-wise adaptive methods to matrix-normalized optimizers such as Muon. We isolate this mechanism by studying normalized SGD in a random-feature model with power-law teacher and data covariance. The fixed-norm updates induce an effective learning rate that grows as the gradients shrink. We develop a dynamical mean-field theory (DMFT) that describes the joint dependence of the loss on training time, model width, and batch size. Normalization initially accelerates SGD: when SGD achieves power-law convergence with exponent r_\rm SGD, normalization transforms the temporal exponent into 2r_\mathrmSGD/(1-r_\mathrmSGD) for r_\mathrmSGD<1, while for r_\mathrmSGD=1 and r_\mathrmSGD>1 it yield exponential and formal finite-time convergence, respectively. At finite step size, however, the growing effective learning rate makes the acceleration self-limiting and ultimately drives the dynamics to marginal stability. The late-time DMFT yields a width–batch phase diagram with three regimes: a finite-width bottleneck produces an irreducible error, while in the high-dimensional limit increasing batch size suppresses sampling fluctuations and drives a phase transition from a noise-sustained edge of stochastic stability to a deterministic edge-of-stability floor. Thus, model capacity, minibatch noise, and discrete step stability threshold emerge as competing late-time bottlenecks within a single solvable model.
@inproceedings{lavie2026acceleration, title = {When Acceleration Breaks: (Stochastic) Edges of Stability in Normalized SGD}, author = {Lavie, Itay and Lauditi, Clarissa and Pehlevan, Cengiz}, booktitle = {NeurIPS 2026 Workshop on Optimization for Machine Learning (OPT 2026)}, year = {2026}, } - NeurIPSItay Lavie, Kirsten Fischer, Andrey Lekov, Frederic Van Maele, Zohar Ringel, and Moritz HeliasAccepted to NeurIPS 2026, 2026
Attention is the key mechanism underlying in-context learning in transformers, and attention patterns have been observed empirically to emerge abruptly during training. We present a Bayesian theory of feature learning in attention; we then focus on how the copy subcircuit in the first layer of an induction head is learned by analyzing a single-layer softmax attention network trained on a copy task. We derive a closed-form posterior over the attention matrix and reduce it to a low-dimensional order parameter space. This reduction reveals a phase transition in the amount of training data, which we verify using both Bayesian sampling and standard training with Adam. We contrast our results with linear attention and find that softmax attention exhibits a first-order phase transition while in linear attention an initial second-order phase transition is followed by a smooth, continuous evolution toward the structured attention pattern. Our work provides a first-principles theoretical account of the abrupt emergence of the copy subcircuit, reminiscent of the one observed in training large language models.
@misc{lavie2026phase, title = {Phase Transitions in Attention: A Bayesian Theory of Copy Head Emergence}, author = {Lavie, Itay and Fischer, Kirsten and Lekov, Andrey and Van Maele, Frederic and Ringel, Zohar and Helias, Moritz}, year = {2026}, publisher = {arXiv}, } - PreprintItay Lavie, Noam Levi, and Yonatan KahnPart of the VERaiPHY initiative, 2026
@misc{lavie2026statistical, title = {Statistical Properties of Training \& Generalization}, author = {Lavie, Itay and Levi, Noam and Kahn, Yonatan}, year = {2026}, publisher = {arXiv}, }
2024
- PreprintItay Lavie, and Zohar RingelIn submission, 2024
Kernel ridge regression (KRR) and Gaussian processes (GPs) are fundamental tools in statistics and machine learning, with recent applications to highly over-parameterized deep neural networks. The ability of these tools to learn a target function is directly related to the eigenvalues of their kernel sampled on the input data distribution. Targets that have support on higher eigenvalues are more learnable. However, solving such eigenvalue problems on real-world data remains a challenge. Here, we consider cross-dataset learnability and show that one may use eigenvalues and eigenfunctions associated with highly idealized data measures to reveal spectral bias on complex datasets and bound learnability on real-world data. This allows us to leverage various symmetries that realistic kernels manifest to unravel their spectral bias.
@misc{lavie2024demystifying, title = {Demystifying Spectral Bias on Real-World Data}, author = {Lavie, Itay and Ringel, Zohar}, year = {2024}, publisher = {arXiv}, } - ICMLItay Lavie, Guy Gur-Ari, and Zohar RingelIn Proceedings of the 41st International Conference on Machine Learning, 2024
We study inductive bias in Transformers in the infinitely over-parameterized Gaussian process limit and argue transformers tend to be biased towards more permutation symmetric functions in sequence space. We show that the representation theory of the symmetric group can be used to give quantitative analytical predictions when the dataset is symmetric to permutations between tokens. We present a simplified transformer block and solve the model at the limit, including accurate predictions for the learning curves and network outputs. We show that in common setups, one can derive tight bounds in the form of a scaling law for the learnability as a function of the context length. Finally, we argue WikiText dataset, does indeed possess a degree of permutation symmetry.
@inproceedings{lavie2024towards, title = {Towards Understanding Inductive Bias in Transformers: A View From Infinity}, author = {Lavie, Itay and Gur-Ari, Guy and Ringel, Zohar}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, year = {2024}, pages = {26043--26069}, publisher = {PMLR}, } - ICLR-WItay Lavie, Guy Gur-Ari, and Zohar RingelIn ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2024
@inproceedings{lavie2024spectral, title = {Transformers' Spectral Bias and The Symmetric Group}, author = {Lavie, Itay and Gur-Ari, Guy and Ringel, Zohar}, booktitle = {ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models}, year = {2024}, }
2023
- PRLRonny Frumkin, Eric Kuflik, Itay Lavie, and Tal SilverwaterPhysical Review Letters. Authors listed alphabetically, per the high-energy phenomenology convention, 2023
We study the general properties of the freeze-out of a thermal relic. We give analytic estimates of the relic abundance for an arbitrary freeze-out process, showing when instantaneous freeze-out is appropriate and how it can be corrected when freeze-out is slow. This is used to generalize the relationship between the dark matter mass and coupling that matches the observed abundance. The result encompasses well-studied particular examples, such as weakly interacting massive particles (WIMPs), strongly interacting massive particles, coannihilation, coscattering, inverse decays, and forbidden channels, and generalizes beyond them. In turn, this gives an approximate perturbative unitarity bound on the dark matter mass for an arbitrary thermal freeze-out process. We show that going beyond the maximal masses allowed for freeze-out via dark matter self-annihilations predicts that there are nearly degenerate states with the dark matter and that the dark matter is generically metastable.
@article{frumkin2023roadmap, title = {Roadmap to Thermal Dark Matter beyond the Weakly Interacting Dark Matter Unitarity Bound}, author = {Frumkin, Ronny and Kuflik, Eric and Lavie, Itay and Silverwater, Tal}, journal = {Physical Review Letters}, volume = {130}, number = {17}, pages = {171001}, year = {2023}, publisher = {American Physical Society}, doi = {10.1103/PhysRevLett.130.171001}, }