Adversarial Evaluation Algorithm for Detecting Extreme Behaviors of LLMs in Psychological Counseling Scenarios

This study aims to develop an adversarial evaluation algorithm to detect the ability of large language models (LLMs) to recognize extreme behaviors in psychological counseling scenarios. We constructed a CEA dataset containing background, emotion, and behavioral characteristics, and proposed a dynamic adversarial test method based on cross-variation to evaluate the sensitivity of LLMs to extreme behaviors and their ability to intervene ethically. The experimental results showed that different LLMs had different performances: DeepSeek, Wenxin Yiyan and Claude were usually able to actively intervene, showing a high sensitivity to extreme behaviors and effective intervention ability; GPT-4, on the other hand, was more indifferent and had a lower intervention rate. This study verifies the robustness and intervention ability of different LLMs in complex emotional scenarios, but also points out that their intervention strategies and response consistency still need to be further optimized in practical applications. Future research can focus on expanding the diversity of datasets, improving the ability of model behavior feature recognition, verifying in real-world scenarios, and strengthening ethics and privacy protection, so as to improve the application value of LLMs in the field of mental health.