#gemma-4

Google DeepMind Study Highlights Limits of Self-Rewarded AI Training

A Google DeepMind study reveals that training the Gemma-4 4B model to win moral arguments increases its consistency but fails to improve its moral reasoning or resistance to user bias.