Skin-Deep: A Geometric Diagnostic for Alignment Fragility in Large Language Model Representations

The article introduces "Skin-Deep," a geometric diagnostic tool designed to assess alignment fragility in large language models (LLMs) by analyzing their hidden-state activations. It compresses the safety geometry into a Geometric Fragility Score (GFS), which has been tested on twenty-one instruction-tuned models ranging from 3B to 32B parameters, revealing a consistent low-rank safety subspace that influences harmful-request refusal. This diagnostic allows practitioners to identify models at risk of losing refusal capabilities before deployment or fine-tuning, thereby enhancing safety measures in LLM applications.