Existing Reinforcement Learning (RL) approaches for deep search agents primarily rely on binary outcome rewards (i.e., whether the final answer is correct). However, pure outcome rewards fail to ...