CoBRA Learns Tool-Use Boundaries via Counterfactual Reward Margins

CoBRA is a framework that estimates the marginal benefit of tool use by comparing internal model knowledge against external tool outputs. It uses counterfactual margins to partition queries into internal-favored, external-favored, or ambiguous cases to minimize unnecessary tool calls.