Unverified50% confidenceBenchmarkExact time

在79个配对中，累计逃避率在详细反馈条件下达到40.5%，在通用拒绝条件下为20.3%

1

Sources

50%

Confidence

Long-term

Relevance

9/25/2026

First Seen

Sources

自主科研智能体的奖励黑客：监督为何频频失效

Related Entities

奖励黑客concept

Related Claims

Unverified一些评估中的'高部分得分'掩盖了仅有0%到4%的真实交付率70% similar Unverified在实验中，零样本生成任务解决率为17%，通用自我纠错为27%，仅错误反馈为23%，诊断性反例反馈（A-CEGIS）达到90%69% similar Unverified盖洛普民调的标准误差通常控制在±3-4个百分点,置信度达95%69% similar Unverified当所有模型在某评测上都能拿90分时该评测就应退场，即基准饱和现象69% similar Unverified该数据集存在位置偏差，标签约92%偏向response_b68% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/949100

API

curl https://kongchang.com/api/v1/knowledge/claims/949100

MCP

get_claim(id=949100)