Unverified50% confidenceBenchmarkExact time
该研究发现同一模型不同部署时间点在代码生成和逻辑推理任务上通过率可能出现10%至20%的显著波动，且方向不总是单向改善
1
Sources
50%
Confidence
Long-term
Relevance
7/23/2026
First Seen
Sources
Related Claims
Unverified早期研究发现同一问题用不同表述方式提问，模型性能差异可高达数十个百分点75% similarUnverifiedStanford University published a 2023 study titled "How Is ChatGPT's Behavior Changing over Time?", experimentally demonstrating that different deployed versions of the same model can show significant fluctuations of 10% to 20% in pass rates on tasks such as code generation and logical reasoning.74% similarUnverifiedResearch shows that the performance gap between different prompt designs can exceed 10 percentage points even for the same model.73% similarUnverifiedResearch shows that high-quality prompts can bring a 20%–40% improvement in task completion rates compared to casually written prompts under the same model conditions.73% similarPartially VerifiedResearch shows that the output quality of the same model can vary by several times under different prompts.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/596168API
curl https://kongchang.com/api/v1/knowledge/claims/596168MCP
get_claim(id=596168)