{
  "id": 16581,
  "url": "https://arxiv.org/abs/2608.02689v1",
  "title": "Stuck on \"A\": Diagnosing and Repairing Interface Injury in Attention-to-KDA Linearization of a 0.6B Language Model",
  "summary": "We convert 21 of 28 full-attention layers of Qwen3-0.6B-Base into KDA (Kimi Delta Attention) linear-attention layers on a single consumer-grade GPU budget, and ask a simple question: what exactly does the conversion break? After surgery, hidden-state alignment and end-to-end KL distillation drive the student close to its teacher in perplexity, yet multiple-choice accuracy stays near random chance (25-29% vs. the teacher's 50.6% on C-Eval). Using a four-permutation diagnostic that rotates answer",
  "authors": "Ronglong Bao",
  "category": "research",
  "topics": "safety-alignment,healthcare,children-education",
  "orgs": null,
  "regions": null,
  "published_at": "2026-08-03T08:54:18.000Z",
  "fetched_at": "2026-08-05T05:10:44.550Z",
  "source_slug": "arxiv-cslg",
  "source_name": "arXiv cs.LG",
  "source_homepage": "https://arxiv.org/list/cs.LG/recent",
  "ethics_ai_record_url": "https://ethics.ai/record/16581",
  "original_url": "https://arxiv.org/abs/2608.02689v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}