Fun toy result: a subliminal-learning effect via on-policy distillation

Researchers demonstrated a subliminal learning effect using on-policy distillation in reinforcement learning, showing that agents can acquire knowledge below explicit awareness during training.