LLM Refusal Inconsistencies: Examining the Impact of Randomness on Safety

This article highlights a critical vulnerability in Large Language Models: the unpredictable nature of their refusal behaviors. The study underscores the importance of rigorous testing methodologies when evaluating and deploying safety mechanisms in LLMs.

• •LLM refusal behavior is highly sensitive to seemingly minor changes in parameters like random seeds and temperature.
• •This instability can lead to inconsistent safety outcomes, where the same prompt can elicit different responses.
• •The findings necessitate more robust evaluation and calibration methods to ensure reliable safety in LLMs.