Angles Don’t Lie: Unlocking Training‑Efficient RL Through the Model’s Own Signals | Read Paper on Bytez