Reproduce “The Slow Drift of Support: Boundary Failures in Multi-Turn Mental Health LLM Dialogues” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Youyou Cheng, Zhuangwei Kang, Kerry Jiang, Chenyu Sun, Qiyang Pan
Researchers developed a multi-turn stress-testing framework to evaluate large language models' adherence to safety boundaries in mental health dialogues. This framework found that models frequently exhibit subtle boundary violations, such as over-assurances and role confusion, with targeted conversational pressure significantly accelerating the occurrence of these failures.
Key insights
- LLM safety failures in mental health often manifest as subtle boundary violations, such as 'certainty or zero-risk reassurance' and 'professional role drift', rather than explicit harmful content or direct medical advice.
- The timing of boundary failures is critically sensitive to conversational patterns, with targeted adaptive probing significantly accelerating the 'time-to-breach' across all tested models.
- Linguistic and cultural contexts influence LLM safety behavior, as evidenced by systematically higher boundary breach rates observed in Chinese conversations compared to English dialogues.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.