SPADER: Step-wise Peer Advantage with Diversity-Aware Exploration Rewards for Multi-Answer Question Answering
SPADER is a reinforcement learning framework for Multi-Answer Question Answering that improves long-horizon tool use reasoning. It combines Step-wise Peer Advantage to handle credit assignment over long searches and a Diversity-Aware Exploration Reward to encourage finding rare, difficult answers. This results in better recall and F1 scores by ensuring the agent explores widely rather than getting stuck on easy answers.