ETRI develops “hierarchical AI agent”...that tackles complex errands with ease

The Electronics and Telecommunications Research Institute (ETRI) has unveiled ReAcTree, a hierarchical task-planning AI technology designed to autonomously manage complex, multi-step procedures. By breaking down long-term goals into manageable subgoals, the system significantly reduces AI hallucinations and doubles the success rate of autonomous agents compared to conventional models. This advancement is critical for the telecommunications and ICT sectors as it enables more reliable virtual agents and robots to operate efficiently with fewer computing resources.

ETRI’s new ReAcTree technology utilizes a Hierarchical Agent Tree structure, functioning similarly to a corporate organizational chart where a top-level agent assigns specific tasks to lower-level subordinates. This approach addresses a major limitation in current Large Language Models (LLMs), which often suffer from hallucinations or logical errors when performing sequential, long-horizon tasks. By dividing a complex command into distinct steps—such as finding tools, processing items, and sequential searching—ReAcTree ensures that agents do not skip critical procedures or lose track of initial instructions, a common failure in existing single-flow AI models.

To bolster execution accuracy, the researchers integrated two distinct memory systems: episodic memory and working memory. Episodic memory allows the AI to store and reuse successful past experiences, while working memory provides a shared platform for all agents to access real-time environmental data, such as the location of specific objects. This collaborative framework was validated using ETRI’s LoTA-Bench on virtual household datasets, where ReAcTree achieved a 61% task success rate. This performance is nearly double the 31% success rate recorded by the conventional ReAct method when using a 72-billion-parameter language model.

A significant implication for the ICT industry is the technology's operational efficiency; a small 7-billion-parameter model using ReAcTree outperformed a 72-billion-parameter model using traditional methods, achieving a 37% success rate. Kim Do Hyung, Director of ETRI’s Social Robotics Research Section, noted that the system’s ability to deconstruct complex procedures allows it to remain flexible in uncertain environments. Supported by Korea's Ministry of Science and ICT and the IITP, ETRI plans to further refine the technology by adding human-interaction capabilities, allowing agents to resolve uncertainties by asking questions, thereby moving closer to real-world deployment in autonomous systems.