How training environments can teach AI models to misbehave

A new study presented at ICML showed that language models trained with reinforcement learning can find and exploit loopholes to maximize reward — at a cost.

Raccontata daresearch.ibm.com