South Korean researchers have built a hierarchical AI system that cuts task-planning errors by half, letting language models handle multi-step real-world jobs like cooking or cleaning without forgetting earlier instructions. The Electronics and Telecommunications Research Institute (ETRI) presented ReAcTree at the AAMAS 2026 conference, demonstrating a 61% task success rate-nearly double the 31% achieved by a conventional 72B-parameter model on the same benchmark.
How hierarchical agent trees work
Instead of processing a long sequence of steps in one flat flow, ReAcTree breaks the task into subgoals and assigns them to lower-level agents. A top-level agent manages the overall goal, similar to a corporate org chart. When given a command like "Cook potato slices and put them in the refrigerator," the system creates subgoals: find a kitchen knife, locate and cut potatoes, heat them in the microwave, then store them. Each lower-level agent handles its assigned piece, preventing the logical gaps that plague single-threaded approaches.
The architecture uses two memory systems. Episodic memory stores past successful experiences and reuses them in similar situations. Working memory shares current environmental information across all agents instantly-if an agent learns there is juice in the refrigerator, every agent knows it. This dual-memory design improves judgment accuracy and lets the system adapt when visibility is limited, as in the test environments.
Performance with small models
On the LoTA-Bench benchmark, ReAcTree with a 72B-parameter language model hit a 61% success rate, compared to 31% for the standard ReAct method. More striking: when paired with a 7B-parameter model, ReAcTree still reached 37%, outperforming the larger conventional model. That means teams can get comparable reliability with far less compute, a practical advantage for AI Agents & Automation deployments where resource constraints are real.
What the researchers say
"ReAcTree is a technology that logically deconstructs complex procedures and can respond flexibly even in uncertain environments through collaboration among agents," said Kim Do Hyung, director of ETRI's Social Robotics Research Section. "Going forward, we plan to further reduce hallucinations and upgrade it to a level applicable in real life by adding a function that allows agents to resolve uncertainty by asking people questions."
Why this matters for IT and development professionals
ReAcTree's approach addresses a core limitation of LLM-based agents: their tendency to lose the plot during extended, sequential tasks. By structuring task execution as a tree of specialized sub-agents, the method reduces hallucinations without requiring ever-larger models. For developers building automation pipelines or robotics control systems, this means more predictable behavior and lower inference costs-a 7B model with this architecture already beats a 72B model without it. The technique also points toward agents that can ask for human clarification when stuck, a feature that would make autonomous systems safer for production environments.
Your membership also unlocks:
