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

arXiv 2603.12740·high signal

Accepted at ICLR 2026, ToolTree replaces greedy reactive tool selection with MCTS using dual-stage LLM evaluation and bidirectional pruning — eliminating unproductive branches both before and after tool execution. Achieves approximately 10% performance improvement over state-of-the-art planning baselines while maintaining computational efficiency across 4 benchmarks covering open-set and closed-set tool planning tasks.