Unverified50% confidenceFactExact time
In open-ended generation tasks, the study observed occasional accuracy drops, which were more pronounced in non-English responses.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Claims
Unverified实验数据显示，经过5-10轮迭代训练后，模型在低频词汇和长尾知识上的覆盖率下降超过40%，而输出文本在表面流畅性上几乎没有可察觉的退化73% similarUnverified研究表明在INT4/INT8精度下多数推理和生成任务的输出质量下降幅度在1-3%以内72% similarUnverifiedAnthropic's research shows that even mild alignment training can cause a 2–8 percentage point decline in model performance on certain reasoning-intensive tasks.71% similarUnverifiedThe study found that models' accuracy on semantically plausible distorted statements is actually higher than their performance on meaningless random replacements.70% similarUnverifiedResearch shows that in certain tasks, simply changing '请判断' to '请分析' in a prompt can cause accuracy fluctuations of 5–10 percentage points.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/902524API
curl https://kongchang.com/api/v1/knowledge/claims/902524MCP
get_claim(id=902524)