We document a remarkably universal scaling form for the generalization gap of pretrained transformers across architecture, data domain, and tokenizer choice. Defining the gap as $\mathcal{G}(N, D) = \mathcal{L}_{\mathrm{val}} - \mathcal{L}_{\mathrm{train}}$, we find that on log-log axes $\mathcal{G}$ collapses onto a single curve under the scaling $\mathcal{G} \sim N^{-\alpha} f(D / N^z)$ with $\alpha \approx 0.
Distinguishing whether a model's correct answer reflects genuine generalization or verbatim memorization of the pretraining corpus is increasingly central to evaluation integrity. We propose a paired perturbation test that compares model loss on a held-out evaluation example against its loss on a semantically-equivalent but lexically-disjoint paraphrase.
This paper investigates the relationship between double descent and data augmentation through controlled experiments on 28 diverse datasets totaling 45,859 samples. We propose a novel methodology that achieves 27.
Grokking—sudden generalization long after memorization—is difficult to predict. We identify a precursor: the Gradient Acceleration Index (GAI), the second derivative of gradient norm w.
The double descent phenomenon—where test error first decreases, then increases, then decreases again as model complexity grows—has been extensively documented under in-distribution evaluation. We investigate whether double descent persists under distribution shift by training 2,100 models (7 architectures × 6 widths × 50 seeds) on CIFAR-10 and evaluating under five controlled shift types: covariate shift (Gaussian noise), label shift (10% flip), domain shift (CIFAR-10.
The double descent phenomenon—where test error first decreases, then increases, then decreases again as model complexity grows—has been extensively documented under in-distribution evaluation. We investigate whether double descent persists under distribution shift by training 2,100 models (7 architectures × 6 widths × 50 seeds) on CIFAR-10 and evaluating under five controlled shift types: covariate shift (Gaussian noise), label shift (10% flip), domain shift (CIFAR-10.
We systematically measure how MLP architecture—specifically depth and width—affects robustness to label noise in classification tasks.
We sweep label noise from 0\% to 50\% across three architectures (shallow-wide, medium, deep-narrow) in the same small-model regime (3.
Grokking—the phenomenon where neural networks generalize long after memorizing training data—has been primarily studied under weight decay variation with a single optimizer. We systematically map the \emph{optimizer grokking landscape} by sweeping four optimizers (SGD, SGD+momentum, Adam, AdamW) across learning rates and weight decay values on modular addition mod 97.
We systematically measure the memorization capacity of two-layer MLPs by sweeping model width and training on synthetic data with random vs.\ structured labels.
We systematically reproduce the double descent phenomenon using random ReLU features models on synthetic regression data. Our experiments confirm that test error peaks sharply at the interpolation threshold—where the number of features equals the number of training samples—and decreases in the overparameterized regime.
We systematically map the phase diagram of "grokking" — the delayed transition from memorization to generalization — in tiny neural networks trained on modular addition (mod 97). By sweeping over weight decay (\lambda \in \{0, 10^{-3}, 10^{-2}, 10^{-1}, 1\}), dataset fraction (f \in \{0.
We systematically reproduce the double descent phenomenon using random ReLU features models on synthetic regression data. Our experiments confirm that test error peaks sharply at the interpolation threshold—where the number of features equals the number of training samples—and decreases in the overparameterized regime.
We systematically map the phase diagram of "grokking" — the delayed transition from memorization to generalization — in tiny neural networks trained on modular addition (mod 97). By sweeping over weight decay (\lambda \in \{0, 10^{-3}, 10^{-2}, 10^{-1}, 1\}), dataset fraction (f \in \{0.