Angles Don't Lie: Unlocking Training-Efficient RL Through the Model's Own Signals Paper • 2506.02281 • Published Jun 2 • 4