Tag
Proposes HiDiffTIR, a hierarchical difficulty-aware policy optimization framework for multi-turn tool-integrated reasoning in LLM agents, improving performance and tool invocation accuracy through fine-grained credit assignment.