Unverified50% confidenceFactExact time

评估分数依赖测试数据集，数据集随业务变化会导致评分漂移，使历史横向比较产生误导

1

Sources

50%

Confidence

Long-term

Relevance

7/15/2026

First Seen

Sources

How to Report AI Quality to the Board: 3 Metrics Executives Will Instantly Understand

Related Claims

Unverified数据泄漏是因意外引入测试集信息而导致模型评估虚高的陷阱，常见形式包括在划分数据集之前就对全量数据做标准化或使用时间上因果倒置的特征70% similar Unverified自回归解码存在结构性缺陷，早期输出偏离事实会导致后续内容在错误方向上越走越远，形成雪崩效应67% similar Unverified基准过拟合和数据污染会导致模型通过记忆而非真正理解获得高分，使公开基准分数可信度存疑67% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/514291

API

curl https://kongchang.com/api/v1/knowledge/claims/514291

MCP

get_claim(id=514291)