Coaching API-calling giant language mannequin (LLM) brokers calls for large quantities of high-quality trajectories. Nonetheless, gathering such information at scale sometimes requires totally applied environments with executable APIs and sensible, pre-populated backend databases, creating a significant bottleneck for scalability. To beat this, we suggest an environment-free artificial information technology strategy that leverages LLMs as on-the-fly digital world fashions. Given solely API specs, our technique generates trajectories mimicking interactions between an agent and a stateful atmosphere. Particularly, an LLM first generates numerous duties solvable with the offered APIs. A instructor agent then iteratively solves every activity whereas an LLM simulator generates coherent artificial API responses conditioned on the duty context and simulation historical past. Lastly, an LLM choose filters the trajectories to make sure the standard of the ensuing dataset. We consider our strategy on the difficult AppWorld and OfficeBench benchmarks, which embody each information-retrieval and state-changing duties. High-quality-tuning fashions on our artificial information yields vital efficiency beneficial properties, demonstrating that efficient supervision for API-calling brokers could be generated with none executable atmosphere. Our outcomes set up LLM-based API simulation as a sensible, scalable answer for coaching brokers throughout numerous API ecosystems.
