Validation-Gated Causal Interventions for Interpreting High-Stakes Large Language Model Behavior: A Case Study in Suicidality Detection
Researchers used a validation-gated framework to rigorously test what features large language models use for binary suicide detection. They found a single, compact, semantic feature in mid-network layers is necessary and low-rank for this task. This feature encodes severity and transfers across different models and datasets, showing how internal representations can be actionable if they are causally necessary.