상황인지 기반 사회적 촉매(Social Catalyst)형 AI 컴패니언 시스템 아키텍처 설계 및 최적 LLM 엔진 선정을 위한 비교 연구

초록

This study addresses the paradox of being “Alone Together” where technological advancement often leads to human alienation,by proposing an approach to design AI as a “Social Catalyst” to promote human-to-human interaction. We designed a systemarchitecture for a “Social Catalyst AI Companion” that actively intervenes in shared media viewing environments based onmultimodal information. The core objective of this research is to identify the engineering mechanisms that enable such interactionfacilitation and to derive the optimal intelligent engine within the designed structure. To this end, a comparative experiment wasconducted on six major Large Language Models (LLMs), combining qualitative indicators such as persona consistency withquantitative indicators such as speech quality and operational efficiency. The experimental results demonstrated a clear trade-offbetween intelligence and speed; Gemini-3 achieved the highest Relative Performance Index (RPI) of 100 in situational awarenessand Transition Relevance Place (TRP) prediction precision, while Llama-4 showed superior real-time responsiveness with a latencyof 3.41s. To overcome the limitations of single LLM engines, we propose a Hybrid Intelligent Architecture based on a“situation-tailored engine switching strategy”. Rather than directly verifying the increase in social interaction, this study issignificant in presenting the systemic conditions and design possibilities for AI to function as a social catalyst, focusing on keymechanisms such as conversation elicitation and intervention timing.

키워드