返回全部动态
政策变更下学习产物的变更溯源监督
原标题：Change-Provenant Supervision: Governing Learned Artifacts Under Policy Change
AI 摘要
论文研究在政策变更下如何对学习产物（如LoRA适配器）进行失效、重建与准入。作者提出将记录的谱系（用于提出影响范围和解释）与独立的当前契约重验证分离，并在Qwen3-14B LoRA包和合成修正测试中验证：仅靠依赖图无法保证安全准入，独立重验证可拒绝陈旧包。
以上摘要由 AI 生成，可能存在误差。事实请以原文为准。
正文节选
Change-Provenant Supervision: Governing Learned Artifacts Under Policy Change Abstract A recorded dependency graph cannot certify that it contains no omitted edge. For learned artifacts, graph-scoped invalidation therefore cannot by itself justify admission after authority changes. The problem persists when an artifact remains behaviorally plausible and output testing misses its provenance-stale derivation. We study how such artifacts can be invalidated, rebuilt, and admitted under versioned ins
发布时间：2026-09-25 12:00
抓取时间：2026-09-25 14:49
来源机构：arXiv