{
  "id": 6001,
  "url": "https://arxiv.org/abs/2604.11838v1",
  "title": "A Layer-wise Analysis of Supervised Fine-Tuning",
  "summary": "While critical for alignment, Supervised Fine-Tuning (SFT) incurs the risk of catastrophic forgetting, yet the layer-wise emergence of instruction-following capabilities remains elusive. We investigate this mechanism via a comprehensive analysis utilizing information-theoretic, geometric, and optimization metrics across model scales (1B-32B). Our experiments reveal a distinct depth-dependent pattern: middle layers (20\\%-80\\%) are stable, whereas final layers exhibit high sensitivity. Leveraging ",
  "authors": "Qinghua Zhao, Xueling Gong, Xinyu Chen, Zhongfeng Kang, Xinlu Li",
  "category": "research",
  "topics": "safety-alignment,finance-investment",
  "orgs": null,
  "regions": null,
  "published_at": "2026-04-12T12:57:12.000Z",
  "fetched_at": "2026-07-14T16:32:11.182Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/6001",
  "original_url": "https://arxiv.org/abs/2604.11838v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}