OrchDAG: Complex Tool Orchestration in Multi-Turn Interactions with Plan DAGs
This work addresses the complexity of multi-turn tool use for AI agents, but it is incremental as it builds on existing agentic tool calling methods.
The paper tackles the problem of complex multi-turn tool interactions in agentic systems by introducing OrchDAG, a synthetic data generation pipeline that models tool execution as directed acyclic graphs with controllable complexity, and shows that the dataset provides a challenging benchmark and the proposed graph-based reward enhances RLVR training when combined with GRPO-style algorithms.
Agentic tool use has gained traction with the rise of agentic tool calling, yet most existing work overlooks the complexity of multi-turn tool interactions. We introduce OrchDAG, a synthetic data generation pipeline that models tool execution as directed acyclic graphs (DAGs) with controllable complexity. Using this dataset, we benchmark model performance and propose a graph-based reward to enhance RLVR training. Experiments show that the dataset presents a challenging but solvable benchmark, and the proposed reward is effective when combined with GRPO-style algorithms, highlighting the importance of leveraging topological structure and data complexity in multi-turn tool use.