Hancom said on the 30th that it had posted two papers on its self-developed core artificial intelligence (AI) technology on arXiv, a world-renowned platform for publishing technical research papers. arXiv is an arena where global big tech corporations such as Google, Meta, and OpenAI share their latest research results fastest before official conference presentations to prove their technological prowess.
At the center of the agentic AI paper released this time (SLM-Based Agentic AI with P–C–G: Optimized for Korean Tool Use) is the "P–C–G (planning–calling–generating) model." This is a new approach that goes beyond the limits of the conventional method in which one massive AI handles everything. Until now, agentic AI that plans on its own and uses tools was widely thought to require a large language model (LLM), which demands enormous expense and resources.
Hancom divided this task into three stages and designed a structure in which small language models (sLMs) specialized for each task collaborate organically like one team. ▲ Planner ▲ Caller ▲ Generator fall into this category. When evaluated with data Hancom built in-house, the system showed performance comparable to or better than large models in complex tasks that call multiple tools in sequence or in identifying whether the tools required for a user request exist. It proved that an efficient collaborative structure, not size, can be the key to AI performance.
Hancom's second paper (Chunk Knowledge Generation Model for Enhanced Information Retrieval: A Multi-task Learning Approach) can boost the performance of retrieval-augmented generation (RAG), a technology that improves the accuracy and reliability of generative AI. It improves accuracy when AI searches for external information. Hancom's researchers had the AI generate "search hints," such as keywords or anticipated questions, by itself at the databases construction stage.
Jeong Ji-hwan, Hancom's chief technology officer (CTO), said, "We plan to sequentially apply the 'P–C–G' model and advanced RAG technology to our AI product lineup, including 'Hancom Assistant' and 'Hancompedia,'" adding, "The release of these papers is an important milestone that proves on the global stage that Hancom is not just a corporations that uses AI, but a technology-leading corporations that presents the future direction of AI technology."