[논문 리뷰] Can Large Language Models Resolve Semantic Discrepancy in Self-Destructive Subcultures? Evidence from Jirai Kei

Peng Wang, Xilin Tao외 6명|arXiv (Cornell University)|2026. 01. 08.

Mental Health via Writing인용 수 0

한 줄 요약

본 논문은 Subcultural Alignment Solver(SAS)라는 세 부분 프레임워크(검색, 정렬 보고서, 문화 정렬)를 제시하여 Jirai Kei 서브컬처의 자기파괴 행위를 탐지하는 LLM을 향상시키고, 베이스라인을 능가하며 미세조정된 모델과 비견되는 성능을 보인다.

ABSTRACT

Self-destructive behaviors are linked to complex psychological states and can be challenging to diagnose. These behaviors may be even harder to identify within subcultural groups due to their unique expressions. As large language models (LLMs) are applied across various fields, some researchers have begun exploring their application for detecting self-destructive behaviors. Motivated by this, we investigate self-destructive behavior detection within subcultures using current LLM-based methods. However, these methods have two main challenges: (1) Knowledge Lag: Subcultural slang evolves rapidly, faster than LLMs' training cycles; and (2) Semantic Misalignment: it is challenging to grasp the specific and nuanced expressions unique to subcultures. To address these issues, we proposed Subcultural Alignment Solver (SAS), a multi-agent framework that incorporates automatic retrieval and subculture alignment, significantly enhancing the performance of LLMs in detecting self-destructive behavior. Our experimental results show that SAS outperforms the current advanced multi-agent framework OWL. Notably, it competes well with fine-tuned LLMs. We hope that SAS will advance the field of self-destructive behavior detection in subcultural contexts and serve as a valuable resource for future researchers.

연구 동기 및 목표

• Jirai Kei와 같은 빠르게 진화하는 서브컬처 내에서 자기파괴적 행동을 탐지할 필요성을 동기 부여한다.
• 지식 지연과 의미적 불일치로 인한 현 LLM 접근법의 한계를 식별한다.
• Subcultural Alignment Solver (SAS)를 제안하여 서브컬처 지식을 자동으로 검색, 보고, 정렬한다.
• 프롬프트 방법과 다중 에이전트 벤치마크에 대한 SAS의 효율성을 시연하고, 미세조정된 LLM과의 비교를 포함한다.

제안 방법

• Subculture Retrieval: 대상 서브컬처에 대한 정보를 자동으로 웹에서 검색하고 k개의 결과를 반환한다.
• Alignment Report Generation: 검색된 결과를 어휘와 맥락을 상세히 설명하는 백그라운드 보고서로 합성한다.
• Culture Alignment Solver: 서브컬처 정렬을 수행하여 입력 내용을 해석하고 태스크 솔버를 통해 최종 레이블을 산출한다.

실험 결과

연구 질문

• RQ1 Can current LLM-based methods understand fast-evolving subcultures and their nuanced expressions?
• RQ2 Does a retrieval-plus-alignment approach reduce semantic gaps between general language use and subcultural intent?
• RQ3 How does SAS compare to prompting, multi-agent, and fine-tuned baselines on Jirai Kei self-destructive behavior detection?

주요 결과

• Prompting methods underperform on subculture-specific tasks.
• Agentic frameworks improve understanding, with Self-Refine showing consistent gains across LLMs.
• OWL’s autonomous tool-use does not achieve strongest performance in this subculture setting.
• SAS achieves state-of-the-art performance among tested models and can rival fine-tuned LLMs without additional fine-tuning.
• Retrieval-language choice affects performance, with English retrieval generally yielding richer results, while reporting language benefits from cultural adaptation.
• SAS generalizes to other subcultures (Menhera, Yami Kawaii, Tenshi Kaiwai) per expert scoring of generated reports.