Unverified50% confidenceFactExact time
Scale AI的研究团队发现，部分模型在被污染的基准上的表现比在新编写的同等难度题目上高出10-15个百分点
1
Sources
50%
Confidence
Long-term
Relevance
8/20/2026
First Seen
Sources
Related Claims
Unverified研究表明即使是部分污染（模型只见过题目而非答案），也可能显著提升得分72% similarUnverified评估模型不能只看智能指标，还要看Token使用效率和工具调用次数，工具调用次数过多会导致上下文污染70% similarUnverifiedResearch from Stanford University shows that optimizing prompt wording alone can improve model performance by 20%–50% on certain tasks.69% similarUnverified研究表明，结构化提示词相比模糊描述可将 AI 输出质量提升 40% 以上69% similarUnverifiedAfter using the AI debugging trio template (complete error message + relevant code snippet + expected behavior), the answer accuracy rate increased from 30% to over 90%, and problems were essentially resolved in a single attempt.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/778884API
curl https://kongchang.com/api/v1/knowledge/claims/778884MCP
get_claim(id=778884)