Abstract
Function calling (FC) empowers large language models (LLMs) and autonomousagents to interface with external tools, a critical capability for solvingcomplex, real-world problems. As this ability becomes increasingly central toadvanced AI systems, the need for high-quality, multi-turn training data todevelop and refine it cannot be overstated. Existing data synthesis methods,such as random environment sampling or multi-agent role-playing, are notpowerful enough to generate high-quality data in real-world environments.Practical challenges come in three folds: targeted model training, isolation oftool architecture, and multi-turn logical dependency. To address thesestructural deficiencies, we present FunReason-MT, a novel data synthesisframework for real-world multi-turn tool use. FunReason-MT resolves thecomplexity barrier in multi-turn FC data by employing 1) Environment-API GraphInteractions to gather varied high-quality trajectories, 2) Advanced Tool-QuerySynthesis to simplify hard query construction, and 3) Guided Iterative Chainfor sophisticated CoT generation. Evaluations on Berkeley Function-CallingLeaderboard (BFCLv3) demonstrate the power of our framework: a 4B model builtupon FunReason-MT generated data achieves state-of-the-art performance amongcomparable-sized models, outperforming most close-source models. Furtherperformance improvements on BFCLv4 confirm that FunReason-MT provides areliable and robust source for agentic learning.