Environment-free Synthetic Data Generation for API-Calling Agents
The paper claims API-agent training data can be generated from API specs alone, without running the actual environment.
The method uses LLMs to create tasks, simulate stateful API responses, and filter completed trajectories with an LLM judge. The authors test it on AppWorld and OfficeBench, including retrieval and state-changing tasks. Fine-tuning on the synthetic trajectories improves performance in their evaluations. HF Daily Papers' note
The method uses LLMs to create tasks, simulate stateful API responses, and filter completed trajectories with an LLM judge. The authors test it on AppWorld and OfficeBench, including retrieval and state-changing tasks. Fine-tuning on the synthetic trajectories improves performance in their evaluations. HF Daily Papers' note
score 5