{
  "id": 6333,
  "url": "https://arxiv.org/abs/2604.04385v5",
  "title": "How Alignment Routes: Localizing, Scaling, and Controlling Policy Circuits in Language Models",
  "summary": "We localize the policy routing mechanism in alignment-trained language models. An intermediate-layer attention gate reads detected content and triggers deeper amplifier heads that boost the signal toward refusal. In smaller models the gate and amplifier are single heads; at larger scale they become bands of heads across adjacent layers. The gate contributes under 1% of output DLA, yet interchange testing (p = 120 detects the same motif in twelve models from six labs (2B to 72B), though specific ",
  "authors": "Gregory N. Frank",
  "category": "research",
  "topics": "regulation,safety-alignment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-04-06T03:20:37.000Z",
  "fetched_at": "2026-07-14T16:32:24.291Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/6333",
  "original_url": "https://arxiv.org/abs/2604.04385v5",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}