Mechanistic interpretability as reward signal for RL training of LLMs — SAE features + GRPO + anti-Goodhart framework