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

Peng Wang, Xilin Tao, Siyi Yao, Jiageng Wu, Yuntao Zou, Zhuotao Tian, Libo Qin, Dagang Li

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.

Introduction

Detecting self-destructive behavior, such as drug overdose or self-harm, on social media is a critical task for mental health intervention. However, niche subcultures—such as Japan’s Jirai Kei ("landmine type")—develop unique linguistic norms to signal identity or conceal intent. These linguistic nuances present significant challenges for standard Large Language Models (LLMs).

The authors identify two primary hurdles:

  1. Knowledge Lag: Subcultural slang (e.g., "Amuka" for self-harm) emerges and changes rapidly, far outpacing the years-long training cycles of foundation models.
  2. Semantic Misalignment: Words that are innocuous in general contexts take on dangerous meanings within a subculture. For example, "taking pills" usually implies health management, but in Jirai Kei, it almost exclusively refers to an intentional overdose.
Comparison of standard prompting, general agentic frameworks, and the proposed SAS framework in handling subcultural semantic discrepancy.

As shown in the figure above, standard prompting often misinterprets subcultural terms literally (e.g., "Jirai Kei" as a physical landmine), while general agentic frameworks may fail to align specific actions like "taking pills" with their subcultural intent of self-destruction.

The Subcultural Alignment Solver (SAS) Framework

To bridge this semantic gap, the authors propose the Subcultural Alignment Solver (SAS), a three-stage multi-agent framework designed to act as a "cultural translator":

  1. Subculture Retrieval: A retrieval agent uses external search tools to gather the latest information, slang definitions, and social norms related to a specific target subculture.
  2. Alignment Report Generation: A report agent synthesizes this raw search data into a comprehensive "Subculture Report," highlighting unique terminology and background context.
  3. Culture Alignment Solver: This core component re-interprets and rewrites the input sentence through the lens of the Subculture Report. Only after this "semantic alignment" does the model provide a final classification (e.g., Drug Overdose, Eating Disorder, or Self-Harming).
Detailed architecture of the SAS multi-agent framework, illustrating the flow from retrieval to final task solving.

Experimental Evaluation

The primary benchmark used is JiraiBench, a bilingual (Chinese and Japanese) dataset specifically designed to evaluate the detection of self-destructive content within the Jirai Kei subculture. The researchers evaluated SAS across five diverse LLMs, including Qwen-2.5-7B, Llama-3.1-8B, and DeepSeek V3.2.

Key Results

  • Superiority over General Agents: SAS consistently outperformed the state-of-the-art multi-agent framework OWL across almost all metrics. For instance, using DeepSeek V3.2, SAS achieved an F1 score of 0.7999 for Eating Disorder detection.
  • Competitive with Fine-tuning: In tests using Qwen-2.5-7B, SAS performed comparably to—and sometimes better than—models specifically fine-tuned on Jirai Kei data, without requiring expensive retraining.
  • The Importance of Alignment: The explicit rewriting step proved critical. By forcing the model to explain what a sentence means within the subculture before classifying it, the framework drastically reduced false negatives.

Discussion and Conclusion

SAS demonstrates that retrieval-based alignment is a powerful alternative to fine-tuning for rapidly evolving linguistic domains. Interestingly, the authors found that searching in English often yielded better results than searching in the subculture's native language, likely due to the richness of aggregate resources like "Aesthetics Wiki."

While SAS shows great promise, limitations remain, including dependency on the indexed web for retrieval and higher latency compared to single-pass models. Future work will focus on handling mixed-context environments where subcultural and general language are intertwined.

Summary updated

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.
BibTeX
@misc{wang2026can,
  author = {Peng Wang and Xilin Tao and Siyi Yao and Jiageng Wu and Yuntao Zou and Zhuotao Tian and Libo Qin and Dagang Li},
  title = {{Can Large Language Models Resolve Semantic Discrepancy in Self-Destructive Subcultures? Evidence from Jirai Kei}},
  year = {2026},
  eprint = {2601.05004},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.05004},
}
Figure