More Data, Worse Decisions? Preference Reversals in Neural Networks under Gram Incompatibility

A collection of newly announced arXiv machine-learning and statistics papers covering compositional reliability, robust optimization, calibration, robust estimation, adversarial training, privacy-preserving estimation, feature selection, and uncertainty quantification. Several papers address security- or safety-relevant themes, including poisoning resistance under local differential privacy, adversarial robustness, uncertainty quantification, and harmful decision reversals, but the document reports research findings rather than a specific disclosed vulnerability or active threat.