ARC Framework Restores Reward Fairness in Open-Ended Agent Interactions

ARC (Advantage Regularization via Conditioning) addresses reward model bias in open-ended interactions where diverse agent behaviors like clarification or progress updates distort relative advantage. The method uses strategy-conditioned rollout grouping and hybrid rewards to ensure optimization favors context-appropriate actions rather than just high-reward interaction styles.