{
  "id": 16212,
  "url": "https://arxiv.org/abs/2607.04510",
  "title": "Transplanting, inverting, and preventing a misalignment persona: method-conditional emergent misalignment in Qwen2.5",
  "summary": "arXiv:2607.04510v2 Announce Type: replace-cross Abstract: Emergent misalignment (EM) --- the broad misbehaviour a language model acquires after fine-tuning on narrow harmful data --- is mediated in Qwen2.5 models by a latent persona direction, and that direction is causal in open weights. Transplanting it into a model that shares only pretraining with its source induces broad EM (2.83 $\\pm$ 0.26\\% misaligned against a random-direction floor of $\\sim$1.1\\%), and ablating a model's own direction r",
  "authors": "Lyndon Drake (University of Oxford), Zandi Eberstadt (University of Oxford)",
  "category": "research",
  "topics": "safety-alignment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-08-05T04:00:00.000Z",
  "fetched_at": "2026-08-05T05:10:44.550Z",
  "source_slug": "arxiv-cscy",
  "source_name": "arXiv cs.CY",
  "source_homepage": "https://arxiv.org/list/cs.CY/recent",
  "ethics_ai_record_url": "https://ethics.ai/record/16212",
  "original_url": "https://arxiv.org/abs/2607.04510",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}