Scrambling Reasoning Steps Makes AI Training Rewards More Reliable

Original: CRoS: Causality Reward on Scrambled CoTs Improves Accuracy and Reasoning Faithfulness

Source: arxiv.org ↗