{
  "id": 4452,
  "url": "https://arxiv.org/abs/2605.12339v1",
  "title": "BSO: Safety Alignment Is Density Ratio Matching",
  "summary": "Aligning language models for both helpfulness and safety typically requires complex pipelines-separate reward and cost models, online reinforcement learning, and primal-dual updates. Recent direct preference optimization approaches simplify training but incorporate safety through ad-hoc modifications such as multi-stage procedures or heuristic margin terms, lacking a principled derivation. We show that the likelihood ratio of the optimal safe policy admits a closed-form decomposition that reduce",
  "authors": "Tien-Phat Nguyen, Truong Nguyen, Thin Nguyen, Duy Minh Ho Nguyen, Ngoc-Thanh Dinh, Trung Le",
  "category": "research",
  "topics": "regulation,safety-alignment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-05-12T16:19:42.000Z",
  "fetched_at": "2026-07-14T16:30:59.238Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/4452",
  "original_url": "https://arxiv.org/abs/2605.12339v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}