Benchmarking Recursive-Collapse Warning Claims Under Matched False-Positive Control
在匹配假阳性控制下对递归崩溃警告声明的基准测试
机构 * Independent Researcher(独立研究者)
专题命中 评测与基准 :LLM(abstract,abstract_cn);分类 cs.LG
AI总结 提出Loopzero基准框架,通过方向性遥测模式(增益G、递归持久性p、多样性δ)在匹配假阳性预算下评估递归系统崩溃警告声明,并报告标准检测器未达到可接受工作点。
Comments 29 pages, 7 figures, 2 tables; supplementary materials: 9 pages, 1 figure, 4 tables. Code, derived data packets, and Lean artifact: https://github.com/davidmullett/loopzero-paper-public (release tag lean-v1.0)