Google DeepMind withholds reward optimization from internal chain-of-thought tokens to keep the reasoning process unfiltered for safety monitoring

Opening sign-in…