Apple AI Updates: April 21, 2026

1. Apple Research: VLM Top Logits Leak Task-Irrelevant Information

Apple. Apple researchers Masha Fedzechkina, Eleonora Gualdoni, Rita Ramos, and Sinead Williamson published a paper investigating information leakage in vision-language models across different representational levels. The work demonstrates that even an easily accessible bottleneck defined by a model’s top logit values can leak task-irrelevant information, exposing potential security and privacy risks in how VLMs surface intermediate representations from the residual stream all the way through to the final logits. Source