{
  "id": 727,
  "url": "https://arxiv.org/abs/2606.22686v2",
  "title": "The Geometry of Refusal: Linear Instability in Safety-Aligned LLMs",
  "summary": "Modern Large Language Models (LLMs) rely on extensive safety alignment, yet the mechanistic basis of refusal remains opaque. In this work, we investigate whether safety compliance is a deep semantic decision or a manipulable linear feature. We introduce Contrastive Logit Steering (CLS), a zero-optimization framework that isolates the \"refusal direction\" by contrasting hidden states derived from safe and unrestricted system prompts. Unlike representation engineering methods that intervene on inte",
  "authors": "Shivam Ratnakar, Kartikeya Vats",
  "category": "research",
  "topics": "regulation,safety-alignment,finance-investment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-06-21T22:04:48.000Z",
  "fetched_at": "2026-07-14T14:14:46.033Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/727",
  "original_url": "https://arxiv.org/abs/2606.22686v2",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}