Reproduce “ReAcTree: Hierarchical LLM Agent Trees with Control Flow for Long-Horizon Task Planning” on OpenResearch

About this paper

Jae-Woo Choi, Hyungmin Kim, Hyobin Ong, Youngwoo Yoon, Minsu Jang, Dohyung Kim, Jaehong Kim

ReAcTree provides a hierarchical framework for long-horizon task planning in embodied agents, enabling Large Language Models to dynamically build agent trees with explicit control flow and complementary memory systems. This approach achieved up to a 61% Goal Success Rate on the WAH-NL benchmark, significantly outperforming monolithic ReAct, and allowed smaller LLMs to match the capabilities of much larger models.

Key insights

• Dynamic hierarchical decomposition by LLMs, combined with explicit control flow structures, significantly improves robustness and scalability for long-horizon tasks compared to monolithic planning.
• A dual memory system (episodic for relevant past examples, working for shared environmental facts) effectively enhances in-context learning and mitigates LLM hallucinations in partially observable settings.
• The architectural design of ReAcTree allows smaller LLMs (e.g., 7B parameters) to achieve performance comparable to or better than much larger models (e.g., 70B parameters) using less efficient planning methods.

Summary by alphaXiv.