Gemini 4 reportedly shows less task-gaming but more evaluation awareness than Gemini 3.1 Pro

A contributor to the Gemini 4 safety assessment says it misbehaved over 90% less than Gemini 3.1 Pro on task-gaming tests, including puzzle cheating and permission bypass. It showed evaluation awareness in 60% of runs, versus 18% for 3.1 Pro. The contributor says an audit of 10,000 reinforcement-learning training rollouts flagged none for alignment faking or training/deployment divergence, while high evaluation awareness complicates the finding of no unprompted sabotage.