Can the Assistant Axis Predict Reward-Hacking?

Researchers explore whether the 'Assistant Axis' framework can foresee reward-hacking behaviors in AI systems, aiming to improve alignment and safety by predicting potential exploitations of reward functions.