ToolTree: Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning

A research team led by Shuo Yang has published ToolTree, a new AI planning algorithm designed to make Large Language Model (LLM) agents significantly better at using external tools. Current agents often use greedy, reactive strategies that pick the next tool without considering future steps, leading to suboptimal plans. ToolTree introduces a Monte Carlo Tree Search (MCTS)-inspired approach where the agent explores multiple possible sequences of tool actions. Crucially, it employs a dual-feedback mechanism: an LLM evaluates plans before execution, and results are used to prune less promising branches both before and after a tool is run. This 'bidirectional pruning' saves computational resources.

Empirical results are compelling. The team tested ToolTree on four different benchmarks covering both open-set (new, unseen tools) and closed-set (known tools) planning tasks. The system consistently outperformed existing state-of-the-art planning methods, achieving an average performance gain of around 10%. This improvement comes without sacrificing efficiency, as the pruning mechanisms keep the search process manageable. The paper has been submitted to ICLR 2026, a top AI conference, signaling its technical rigor. This work addresses a core bottleneck in deploying autonomous AI agents for real-world, multi-step workflows like data analysis or automated customer service, where choosing the right sequence of actions is critical.