High‑quality multi‑turn tool‑use data is crucial for training agentic models, yet most synthesis pipelines overlook argument‑level dependencies. Consequently, a model may pick the right tool but still fail because the supplied arguments are fabricated, stale, or weakly grounded. To remedy this, we introduce State‑Guided Data Synthesis with Argument Provenance (SAP). SAP builds reliable tool‑use trajectories through three components: (1) State guidance – the environment state is explicitly tracked and used as the source for subsequent arguments; (2) Argument provenance constraints – each tool argument must be traceable to a concrete entity or value in a prior state, preventing bogus or outdated inputs; (3) Turn‑level validation – after generating each step the whole trajectory is checked for consistency and executability, ensuring long‑range dependencies are satisfied. Using data generated by SAP we train a 4 B‑parameter model, SAP‑4B, which matches or exceeds much larger models on several benchmarks. The source code, synthesized data, and trained weights are released at https://github.com/Zichen1024/SAP.
Review: By coupling state guidance with argument provenance, SAP markedly improves the fidelity and executability of multi‑turn tool‑use datasets, offering an efficient and reproducible pipeline for training large language models in tool interaction scenarios.