{
  "id": 14522,
  "url": "https://arxiv.org/abs/2607.26654",
  "title": "Constitutional Midtraining: Content Presence Drives Alignment Gains",
  "summary": "arXiv:2607.26654v1 Announce Type: cross Abstract: 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 (curricu",
  "authors": "Desiree Cho, Cameron Tice, Bernie Hogan, Hunar Batra, Puria Radmard, Jun Zhao, Nigel Shadbolt",
  "category": "research",
  "topics": "safety-alignment",
  "orgs": "anthropic",
  "regions": null,
  "published_at": "2026-07-30T04:00:00.000Z",
  "fetched_at": "2026-07-30T05:10:24.387Z",
  "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/14522",
  "original_url": "https://arxiv.org/abs/2607.26654",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}