Model Collapse
A machine learning phenomenon in which an AI model, trained on data containing large volumes of AI-generated content, progressively diverges from the true distribution of human-produced data, resulting in degraded output quality, reduced diversity, and increased homogenization of generated content.
Core Facts
Model collapse refers to the phenomenon whereby, after an AI model is trained on data generated by a previous-generation model, the output distribution gradually deviates from the original real data distribution across multiple generations of iteration, losing information in the tail of the distribution.
The concept of Model Collapse was formally proposed in 2023 by research teams from the University of Oxford and the University of Cambridge in a Nature paper.
During model collapse, the model gradually loses the long-tail portions of the data distribution and ultimately converges to monotonous, repetitive, and diversity-lacking outputs.
Timeline (last 90 days)
The mechanism of model collapse is distribution drift caused by recursive training, a learning-level problem; whereas the spread of false claims through document copying is content-level contamination, and the two have different intervention methods.
During model collapse, the model gradually loses the long-tail portions of the data distribution and ultimately converges to monotonous, repetitive, and diversity-lacking outputs.
Model collapse refers to the phenomenon whereby, after an AI model is trained on data generated by a previous-generation model, the output distribution gradually deviates from the original real data distribution across multiple generations of iteration, losing information in the tail of the distribution.
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.
In 2023, researchers from Oxford University and other institutions published a study finding that after only 5–10 generations of recursive training, the degradation in model output quality became visible to the naked eye.
All Facts (7)
Model collapse refers to the phenomenon whereby, after an AI model is trained on data generated by a previous-generation model, the output distribution gradually deviates from the original real data distribution across multiple generations of iteration, losing information in the tail of the distribution.
80%VerifiedThe concept of Model Collapse was formally proposed in 2023 by research teams from the University of Oxford and the University of Cambridge in a Nature paper.
80%VerifiedDuring model collapse, the model gradually loses the long-tail portions of the data distribution and ultimately converges to monotonous, repetitive, and diversity-lacking outputs.
65%Unverified当生成模型反复在自身生成的数据上训练时，输出分布会逐渐退化，低概率事件被系统性遗忘
85%UnverifiedIn 2023, researchers from Oxford University and other institutions published a study finding that after only 5–10 generations of recursive training, the degradation in model output quality became visible to the naked eye.
60%UnverifiedThe mechanism of model collapse is distribution drift caused by recursive training, a learning-level problem; whereas the spread of false claims through document copying is content-level contamination, and the two have different intervention methods.
50%UnverifiedThe capability degradation phenomenon (also known as 'model collapse') has been observed multiple times when fine-tuning with synthetic data generated by the model itself.
50%