Research map / Multi-agent systems and games / LLM agents

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

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ToolTree, a training-free MCTS-based tool planner that uses LLM-judge pre-execution and post-execution scores for bidirectional pruning, evaluated on GTA, m&m, ToolBench and RestBench.

Problem

Greedy or search-based LLM tool planning either lacks foresight about inter-tool dependencies or wastes compute on costly, poorly grounded branch exploration.

Approach

Integrates a pre-execution LLM score into the UCT selection prior and a post-execution LLM-judged utility as the MCTS reward, with pre- and post-pruning thresholds, so search is guided by both foresight and executed-tool feedback without training.

Key results

• On GTA with GPT-4o, ToolTree reaches 66.95 average F1, versus 64.78 for LATS (the strongest baseline).
• On ToolBench with GPT-4o, ToolTree reaches 69.04 average (pass and win), about +2.5 over the strongest baseline.
• On RestBench-TMDB with GPT-4o, ToolTree reaches 74.50 average, about +3.1 over the next best.

Surprising findings

• Removing post-evaluation causes the largest accuracy drop (over 7 points on GTA), and the paper reports that overly conservative judge decisions (false negatives) harm more than permissive ones.
• Using LLaMA-3.3-70B as both internal judge and benchmark evaluator degrades pass rate (38.11 on ToolBench) relative to cross-vendor setups.

Limitations

• Relies on LLM-judge scores for pre- and post-evaluation, with judge error rates of 25.8% (GPT-4o) and 39.4% (GPT-4o-mini) on ToolBench.
• Results on GTA and m&m are reported with GPT-4o and GPT-4o-mini only; the paper does not report dollar cost of API calls.
• Several sections contain inconsistencies (e.g., claimed 10% gains vs. smaller reported margins; described hyperparameter stopping rules), and the cited figures are not fully described in the text.

Experimental setup

Datasets: GTA m&m ToolBench RestBench

Benchmarks: GTA m&m ToolBench RestBench

Metrics: Tool F1 Arg F1 Plan F1 Exec F1 Pass rate Win rate

Baselines: Zero-shot ReAct Chain-of-Thought Best-First search Tree-of-Thought A* Search (ToolChain*) LATS DFSDT Monte Carlo tree search

Builds on: Monte Carlo Tree Search ToolChain* LATS ReAct Tree-of-Thought

Closest papers by meaning

• ToolGen: Unified Tool Retrieval and Calling via Generation (ICLR 2025 · similarity 0.89)
• Tool-Planner: Task Planning with Clusters across Multiple Tools (ICLR 2025 · similarity 0.89)
• Tree-Planner: Efficient Close-loop Task Planning with Large Language Models (ICLR 2024 · similarity 0.89)
• AutoTool: Efficient Tool Selection for Large Language Model Agents (AAAI 2026 · similarity 0.89)
• T1: A Tool-Oriented Conversational Dataset for Multi-Turn Agentic Planning (NeurIPS 2025 · similarity 0.88)
• Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.88)