A Validation-Gated Mechanistic Account of Suicidality Detection in LLMs

The article presents a validation-gated framework for assessing the mechanistic understanding of suicidality detection in large language models (LLMs), specifically using Llama-3.1-8B-Instruct as a case study. It demonstrates that while smaller models can represent suicidality, only larger models effectively utilize this representation in binary suicide detection tasks, revealing a mid-network feature that is semantically linked to suicidality rather than merely keyword-based. This research is significant for practitioners as it highlights the importance of model size and internal feature interpretation in developing reliable mental health applications using LLMs.