{
  "id": 14581,
  "url": "https://arxiv.org/abs/2607.26654v1",
  "title": "Constitutional Midtraining: Content Presence Drives Alignment Gains",
  "summary": "Post-training alignment is often shallow, eroding under fine-tuning. Whether midtraining interventions, cleanly isolated from post-training, can produce durable alignment remains untested. We test this via constitutional midtraining: inserting principled, values-based content into midtraining against a replay-only control at 120B scale. Our 394M-token constitutional corpus, built from Anthropic's Constitution, uses a 2x2 factorial design (curriculum ordering x deliberative reasoning) to produce",
  "authors": "Desiree Cho, Cameron Tice, Bernie Hogan, Hunar Batra, Puria Radmard, Jun Zhao et al.",
  "category": "research",
  "topics": "safety-alignment",
  "orgs": "anthropic",
  "regions": null,
  "published_at": "2026-07-29T09:15:40.000Z",
  "fetched_at": "2026-07-30T05:10:24.387Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/14581",
  "original_url": "https://arxiv.org/abs/2607.26654v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}