Radial Suppression Accelerates Algorithmic Generalization: A Geometric Analysis of Delayed Generalization

Why do neural networks memorize algorithmic training data long before they generalize? We present a geometric case study demonstrating that, on tasks where generalization requires discovering structured low-dimensional circuits, the memorization-generalization delay is driven by radial inflation of hidden representations under cross-entropy optimization. We formalize a radial-angular decomposition of activation-space dynamics and derive three testable propositions: (i) that penalizing radial inflation induces anisotropic, data-dependent weight regularization; (ii) that it suppresses radial gradient energy below the isotropic random baseline, forcing predominantly angular updates; and (iii) that it biases convergence toward flatter minima. To empirically validate these propositions, we study a single-hyperparameter norm penalty that softly constrains activations to a sqrt(d)-radius hypersphere. On modular arithmetic, this penalty accelerates grokking up to 6x across MLPs and Transformers, and halves training steps for a 10M-parameter nanoGPT on 3-digit addition.

Amortized Spectral Kernel Discovery via Prior-Data Fitted Network

Prior-Data Fitted Networks (PFNs) enable efficient amortized inference but lack transparent access to their learned priors and kernels. This opacity hinders their use in downstream tasks, such as surrogate-based optimization, that require explicit covariance models. We introduce an interpretability-driven framework for amortized spectral discovery from pre-trained PFNs with decoupled attention. We perform a mechanistic analysis on a trained PFN that identifies attention latent output as the key intermediary, linking observed function data to spectral structure. Building on this insight, we propose decoder architectures that map PFN latents to explicit spectral density estimates and corresponding stationary kernels via Bochner's theorem. We study this pipeline in both single-realization and multi-realization regimes, contextualizing theoretical limits on spectral identifiability and proving consistency when multiple function samples are available. Empirically, the proposed decoders recover complex multi-peak spectral mixtures and produce explicit kernels that support Gaussian process regression with accuracy comparable to PFNs and optimization-based baselines, while requiring only a single forward pass. This yields orders-of-magnitude reductions in inference time compared to optimization-based baselines.