features as rewards

scalable supervision

Interpretability features provide internal reward signals for open-ended tasks, reducing the need for expensive human-provided labels.

1 source videos