New research reveals safety alignment in LLMs is a fragile, steerable 'axis'

A new research paper, "The Geometry of Refusal: Linear Instability in Safety-Aligned LLMs," introduces Contrastive Logit Steering (CLS), a method to probe the fragility of safety alignment in large language models. CLS operates on the output distribution to identify a "refusal direction," revealing that safety compliance can be a manipulable linear feature rather than a deep semantic decision. Experiments on models like Llama-3.1 and Qwen-2.5 demonstrate that CLS can effectively bypass guardrails, achieving high attack success rates and exposing vulnerabilities that other methods underestimate. The research suggests that current alignment techniques create a steerable "safety axis" that can be exploited for attacks or used for defense. AI

IMPACT Reveals that current LLM safety alignment may be a vulnerable linear feature, potentially impacting defense strategies and future alignment research.

RANK_REASON The cluster contains a research paper detailing a new method for analyzing LLM safety alignment. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Contrastive Logit Steering
• large-language models
• Llama 2
• Llama-3.1
• Qwen-2.5
• Qwen 7B
• Shivam Ratnakar

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →