RewardSpy: debugger detects reward function exploitation in RL training

View on Reddit

A new library wraps reward functions to detect reward hacking during training. It alerts when the policy is exploiting the reward function rather than genuinely improving.