A focused look at this week’s most significant advances in alignment & safety \u2014 4 papers surveyed from arXiv and leading AI labs.<\/p>\n
—<\/p>\n
Safety research is broadening from alignment to encompass robustness, interpretability, and the systemic risks of deployed AI. This week brings new attacks, new defenses, and deeper understanding of model internals.<\/p>\n
RedVox: Safety and Fairness Gaps in Speech Models Across Languages<\/strong> \u2014 Beatrice Savoldi, Sara Papi, Wafa Aissa, Matteo Negri, Luisa Bentivogli<\/em><\/p>\n Speech-capable models are increasingly deployed in real-world applications across languages. Yet their safety and fairness beyond English settings and under naturalistic conditions remain understudied…<\/p>\n arXiv<\/a><\/p>\n The Role of Input Dimensionality in the Emergence and Targeted Control of Adversarial Examples<\/strong> \u2014 Nasrin Malekzadeh Goradel, Niccolo Pancino, Yaser Gholizade Atani, Benedetta Tondi, Giovanni Bellettini, Mauro Barni<\/em><\/p>\n Several theoretical works have tried to explain the adversarial vulnerability of deep neural networks through properties of high-dimensional geometry. However, the assumptions underlying these works a…<\/p>\n arXiv<\/a><\/p>\n Effective Covariance Dynamics in Solvable High-Dimensional GANs<\/strong> \u2014 Andrew Bond, Zafer Do\u011fan<\/em><\/p>\n We study a solvable high-dimensional model of generative adversarial network (GAN) training in which a linear generator learns a low-dimensional subspace from data with structured latent covariance. P…<\/p>\n