Unverified50% confidenceFactExact time
The capability degradation phenomenon (also known as 'model collapse') has been observed multiple times when fine-tuning with synthetic data generated by the model itself.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
Unverified2023年牛津大学Ilia Shumailov团队在Nature上发表研究系统阐述了模型崩溃效应，指出模型在自身生成数据上迭代训练相当于对原始数据分布有损压缩74% similarVerifiedThe concept of Model Collapse was first systematically articulated by research teams from Oxford University and Cambridge University in the 2023 paper The Curse of Recursion: Training on Generated Data Makes Models Forget.71% similarUnverifiedEvaluation highly overlaps with scenarios such as synthetic data generation and model distillation70% similarUnverifiedSynthetic data used to train large language models carries the risk of 'Model Collapse'.70% similarUnverifiedThe term Mode Collapse originates from generative adversarial network research and refers to the tendency of generative models to produce only a small number of safe output patterns.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/831916API
curl https://kongchang.com/api/v1/knowledge/claims/831916MCP
get_claim(id=831916)