2504.03601

This paper introduces APIGen-MT, a novel two-phase framework designed to generate high-quality, verifiable, and diverse multi-turn interaction data for training AI agents. The first phase focuses on creating detailed task blueprints with validated ground-truth actions, utilizing an agentic pipeline and LLM review committees with feedback loops. The second phase transforms these blueprints into realistic conversational trajectories through simulated human-agent interplay. This approach leads to the development of xLAM-2-fc-r models, which demonstrate superior performance on benchmarks like BFCL v3 and τ-bench, especially in multi-turn settings, and exhibit enhanced consistency compared to other frontier models. The project aims to open-source its generated data and models to foster further AI agent research.