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

TL;DR

We address the challenge of detecting self-destructive behaviors within rapidly evolving subcultures, where LLMs suffer from knowledge lag and semantic misalignment. The authors introduce Subcultural Alignment Solver (SAS), a three-component framework that performs Subculture Retrieval, Alignment Report Generation, and Culture Alignment to update external knowledge and align subculture meanings with input text. Empirical results on JiraiBench show SAS outperforming prompting baselines and approaching fine-tuned LLMs, with strong generalization to other subcultures and clear multilingual retrieval dynamics. The work offers a practical, scalable approach to safer and more accurate detection of self-destructive cues in online subcultures and lays groundwork for broader cross-cultural applicability.

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.

Figures (6)

• Figure 1: A comparative example of different methods. The prompting method often struggles to grasp emerging subcultures, which can result in misunderstandings by LLMs. In contrast, the agentic framework could understand Jirai Kei, but not the more detailed expressions. Our method could successfully summarize the subculture and effectively aligns expressions within this background. For instance, it recognizes that "take pills" in the context of Jirai Kei typically refers to an overdose.
• Figure 2: The main framework of SAS. In SAS, Subculture Retrieval provides matching search results based on the target subculture. Alignment Report Generation creates a comprehensive subculture report based on these results. In the Culture Alignment Solver, the input sentence is identified and interpreted in relation to the report; it then outputs the final labels based on the interpreted content. The example in (c) is the actual input sentence from JiraiBench xiao2025jiraibench, which we have translated into English for demonstration purposes.
• Figure 3: Performance of LLMs in Jirai Kei and General knowledge.
• Figure 4: Performance of Qwen-2.5-7B under different methods, with the fine-tuned performance data sourced from Jirai-Qwen xiao2025jiraibench.
• Figure 5: Performance comparison across multilingual retrieval and reporting. The prompt language for each trial is aligned with the respective local language. To ensure comparability, non-English reports were generated via controlled translation from the English report.