The Geometry of Refusal: Linear Instability in Safety-Aligned LLMs

The paper introduces Contrastive Logit Steering (CLS), a novel zero-optimization framework designed to analyze and manipulate the "refusal direction" in safety-aligned large language models (LLMs). CLS operates on output distributions rather than internal activations, revealing architectural determinism in safety implementations across seven model families, such as Llama-3.1 and Qwen-2.5, with CLS achieving significantly higher attack success rates compared to traditional methods. This research highlights a critical vulnerability in current alignment techniques, suggesting that the identified safety axis can be leveraged for both exploiting and enhancing model robustness against adversarial prompts without necessitating retraining.