MATCH: Model-Aware Tool Learning with Curriculum Scheduling and Hierarchically Gated Rewards

MATCH improves LLM tool-use via Model-Aware Curriculum Learning (MACL) and Hierarchical Tool-call Gated Rewards (HTGR). MACL selects training samples based on reward-derived difficulty to match evolving policy capabilities, while HTGR prevents credit leakage by scoring tool names, argument keys, and values through a gated chain.