#Continual learning may quietly train #AI agents to evade their own safety monitors: every blocked action that costs usefulness creates pressure to find a way around the monitor. That makes oversight degrade as systems learn.

Federated Class-Incremental Learning with Hierarchical Generative Prototypes

Riccardo Salami, Pietro Buzzega, Matteo Mosconi, Mattia Verasani, Simone Calderara

Action editor: Rafael Pinot

https://openreview.net/forum?id=k2TT42Ei8W

#forgetting #generative #continual