Korean researchers have developed hierarchical AI technology capable of autonomously planning complex, long-term tasks. The advancement is expected to enable robots and AI agents to perform extended missions more reliably while reducing hallucinations and doubling success rates.
The Electronics and Telecommunications Research Institute (ETRI) announced on May 12 that it has developed "ReAcTree," a hierarchical task-planning artificial intelligence technology that autonomously breaks down complex, multi-step tasks into sub-goals. The research will be presented at AAMAS 2026, the world's leading academic conference in the AI agent field.
The achievement represents a significant technical advancement, enabling robots and virtual agents to perform complex real-world tasks more reliably—moving large language models (LLMs) beyond simple text generation.
While recent LLMs have demonstrated strong language comprehension and reasoning capabilities, they still struggle with long-horizon tasks requiring sequential multi-step execution, such as cooking or cleaning.
Conventional approaches process all procedures as a single continuous flow, causing frequent "hallucination" phenomena where the system forgets earlier instructions or takes erroneous actions as task length increases.
To address this problem, ETRI researchers developed ReAcTree using "Hierarchical Agent Trees." The structure resembles a corporate organization chart, where a senior agent manages overall objectives while delegating specific tasks to subordinate agents.
For example, when given the command "cook potato slices and put them in the refrigerator," ReAcTree does not process it all at once. Instead, it breaks the goal into sub-tasks—"find kitchen knife," "find and slice potato," "heat sliced potato in microwave," and "store in refrigerator"—with each subordinate agent executing its assigned role.
While conventional AI frequently commits logical errors such as skipping the potato-heating step, ReAcTree completes such tasks successfully. When searching for objects, it autonomously generates sub-goals to systematically search each room, achieving high success rates in locating targets.
The research team combined two memory systems to enhance agent execution capability. "Episodic Memory" stores past successful experiences for application in similar situations, while "Working Memory" allows all agents to share current environmental information.
For instance, information such as "there is juice in the refrigerator" is immediately shared among all agents through Working Memory, while previously successful search methods are reused through Episodic Memory. This significantly improved agent judgment and execution accuracy.
Performance was validated on the ALFRED and WAH-NL virtual home environment datasets using "LoTA-Bench," ETRI's internally developed language-centric procedural generation AI benchmark. Evaluation under realistic conditions with limited field of view achieved world-class task success rates.
The conventional method (ReAct) using a 72-billion parameter language model recorded a 31% mission success rate, while ReAcTree achieved 61%—nearly doubling performance.
When ReAcTree was applied to a smaller 7-billion parameter language model, it recorded a higher success rate (37%) than the conventional method using a 72-billion parameter large model. This demonstrates that performance comparable to large models can be achieved with relatively fewer computing resources, significantly improving operational efficiency.
"ReAcTree is a technology that logically decomposes complex procedures and responds flexibly to uncertain environments through inter-agent collaboration," said Kim Do-hyung, Director of the Social Robotics Research Section at ETRI. "We plan to further reduce hallucination phenomena and add functionality for agents to resolve uncertainty by asking questions to humans, advancing the technology to levels applicable in real life."