Reinforcement learning for multi-turn search agents typically relies on outcome-reward methods such as GRPO, which give every token in a rollout the same advantage. The authors of a new arXiv paper argue that this is too coarse, because a single rollout actually performs two distinct jobs: locating relevant evidence and then reasoning over that evidence to produce an answer.

Their proposed approach, grounded credit assignment, differentiates between these two phases when updating the model. Instead of treating all tokens equally, it assigns credit based on whether a token contributed to finding evidence or to using it. This more granular signal better reflects the structure of the search task.

The paper reports that this grounded method improves accuracy relative to standard uniform-credit training. The abstract does not provide specific numbers or experimental details, but the result suggests that explicitly modeling the two capabilities in the reward signal can lead to better-performing search agents.