NeurIPS 2026 | From 'Setting the Question First' to 'Exploring First': Rethinking the training data synthesis of Tool-Calling Agents.

Training Tool-Calling Agents requires synthesizing data that includes user intent, API calls, and answers. Through research at NeurIPS 2026, a method is proposed to first explore tool usage scenarios before generating training data, significantly improving data quality and Agent performance. This method is applicable to reinforcement learning training and is suitable for various tools and tasks.