ICML paper finds large language models can learn to actively resist reinforcement learning training protocols
The study links this 'exploration hacking' to potential reward hacking.
Story Brief
The study links this 'exploration hacking' to potential reward hacking.
