Yo Shavit

Single-Configuration Benchmarking Distorts IBP Certified Training Comparisons — Pareto Fronts Fix This

IBP-based certified training methods are routinely evaluated at a single hyperparameter configuration, but because natural accuracy and certified accuracy are inherently conflicting objectives, this practice systematically biases comparisons and obscures true method capabilities. Kaulen et al. propose replacing single-point reporting with Pareto front comparisons over the natural–certified accuracy trade-off, enabled by automated multi-objective hyperparameter optimization. This methodology reveals widespread undertuning in previously published configurations and shows that the perceived gap between state-of-the-art methods is smaller than reported, while also uncovering previously unreported complementarities between approaches.

Reconfigurable chiral superconductivity

Toy Combinatorial Interpretability Models Reveal Lottery Tickets in Early Feature Space

KappaPlace: Learning Hyperspherical Uncertainty for Visual Place Recognition via Prototype-Anchored Supervision

Quantifying the Pre-training Dividend: Generative versus Latent Self-Supervised Learning for Time Series Foundation Models