{
  "id": 14957,
  "url": "https://arxiv.org/abs/2607.27910v1",
  "title": "A Cross-Architecture Audit of Direction-Based Inference-Time Defences in Vision-Language Models",
  "summary": "Inference time defences against vision language model jailbreaks often subtract a calibrated direction from the residual stream at a chosen decoder layer. We compare five defence candidates across 15 model and layer cells from four architectural families under a magnitude controlled protocol that matches the intervention size for each prompt and pairs every direction with a random control of the same norm. The candidates are the mean image conditioning shift, a CMRM style refusal direction, a Sh",
  "authors": "Xiangyu Yin, Tora Bodin, Rohan Menon, Chih-Hong Cheng",
  "category": "research",
  "topics": "safety-alignment,military-security,transparency",
  "orgs": null,
  "regions": null,
  "published_at": "2026-07-30T09:24:27.000Z",
  "fetched_at": "2026-07-31T05:10:57.675Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/14957",
  "original_url": "https://arxiv.org/abs/2607.27910v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}