评估语义识别：基准效应与排名的稳定性分析

AI 摘要

该论文提出“评估语义识别”概念，研究评估结论是否在声明的语义族内保持不变。通过217条轨迹的实证分析，发现冻结的对比处理在特定语义族内产生零效应，而联合处理则无法识别效应，表明排名依赖规范。作者将分歧归因于单侧终端删除，并提供了可复用的族索引审计方法。

正文节选

Stable Within, Unidentified Across: Semantic Identification of Benchmark Effects and Rankings Abstract Evaluation conclusions depend on evaluator-controlled semantics: legal references, scoreability, and aggregation. We call an artifact-defined endpoint evaluation-semantically identified when it is invariant over a declared family. A frozen 217-row analysis appears stable within its restricted contract family. In TraceElephant, yields precise task-disjoint adoption effects from to , whereas both

发布时间：2026-08-21 12:00

抓取时间：2026-08-21 14:24

来源机构：arXiv