Unverified50% confidenceBenchmarkExact time
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
1
Sources
50%
Confidence
Long-term
Relevance
7/7/2026
First Seen
Sources
Regression to the Mean: How LLMs Are Quietly Killing Innovation and Diversity
hackernewshackernews7/6/2026
Related Claims
Unverified研究表明在超过8-10轮对话后，模型对初始系统提示中规则的遵循率可下降20%-40%79% similarUnverified牛津大学2023年的一项研究量化了模型坍缩效应，仅经过五代自我训练，文本多样性指标下降了约40%74% 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.73% similarUnverifiedResearch indicates that when the context exceeds a certain length, the model's recall accuracy for content in the middle portion may drop below 50%.71% similarUnverifiedWhen a model references five samples from an author and imitates their style, the detection capability of AI text detectors drops significantly.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/170656API
curl https://kongchang.com/api/v1/knowledge/claims/170656MCP
get_claim(id=170656)