研究發現，如果 AI 模型反覆用 AI 生成嘅數據訓練，會出現「模型崩塌」，罕見模式會逐漸消失，模型對現實嘅描述變得愈來愈狹窄。[1][4] 遞歸式訓練會放大抽樣偏差：高機率模式愈來愈多，分佈尾端嘅罕見事件逐代被削弱，最終甚至完全消失。[1][9] 分析顯示，只要加入極少量真實世界數據，或者加入先驗知識限制模型，就有機會阻止模型崩塌，即使合成數據佔絕大多數。[7][33]
What does the new study on AI model collapse find about preventing degradation when models are trained on synthetic data, why does recursiveRecursive training on AI‑generated data can gradually erase rare patterns from a model’s learned distribution, a phenomenon researchers call model collapse.
AI 提示
Create a landscape editorial hero image for this Studio Global article: What does the new study on AI model collapse find about preventing degradation when models are trained on synthetic data, why does recursive. Article summary: The study describes model collapse as a failure mode where recursively trained generative models lose information about the original data distribution, especially its rare or low-probability regions.. Topic tags: general, government, education, academic, general web. Reference image context from search candidates: Reference image 1: visual subject "However, as AI-generated data increasingly populates the internet, an important question arises: What happens when new AI models are trained on datasets containing their previous o" source context "Avoiding Model Collapse in AI Training - Risk Insight" Reference image 2: visual subject "Artificial intelligence models
openai.com
生成式 AI（例如大型語言模型）近年愈來愈多依賴合成數據（synthetic data）訓練——即係由較早一代 AI 產生嘅內容。但研究指出，如果呢個過程持續重複，可能會出現一種問題：模型崩塌（model collapse）。
所謂模型崩塌，係指模型逐漸失去對原始數據多樣性嘅理解，特別係罕見或不尋常嘅模式。當訓練循環反覆依賴 AI 生成嘅內容，而唔係真實世界數據時，呢啲模式會慢慢消失，最後令模型對現實嘅表示出現扭曲。
研究發現，如果 AI 模型反覆用 AI 生成嘅數據訓練，會出現「模型崩塌」，罕見模式會逐漸消失，模型對現實嘅描述變得愈來愈狹窄。[1][4] 遞歸式訓練會放大抽樣偏差：高機率模式愈來愈多，分佈尾端嘅罕見事件逐代被削弱，最終甚至完全消失。[1][9] 分析顯示，只要加入極少量真實世界數據，或者加入先驗知識限制模型，就有機會阻止模型崩塌，即使合成數據佔絕大多數。[7][33]
What does the new study on AI model collapse find about preventing degradation when models are trained on synthetic data, why does recursiveRecursive training on AI‑generated data can gradually erase rare patterns from a model’s learned distribution, a phenomenon researchers call model collapse.
AI 提示
Create a landscape editorial hero image for this Studio Global article: What does the new study on AI model collapse find about preventing degradation when models are trained on synthetic data, why does recursive. Article summary: The study describes model collapse as a failure mode where recursively trained generative models lose information about the original data distribution, especially its rare or low-probability regions.. Topic tags: general, government, education, academic, general web. Reference image context from search candidates: Reference image 1: visual subject "However, as AI-generated data increasingly populates the internet, an important question arises: What happens when new AI models are trained on datasets containing their previous o" source context "Avoiding Model Collapse in AI Training - Risk Insight" Reference image 2: visual subject "Artificial intelligence models
openai.com
生成式 AI（例如大型語言模型）近年愈來愈多依賴合成數據（synthetic data）訓練——即係由較早一代 AI 產生嘅內容。但研究指出，如果呢個過程持續重複，可能會出現一種問題：模型崩塌（model collapse）。
所謂模型崩塌，係指模型逐漸失去對原始數據多樣性嘅理解，特別係罕見或不尋常嘅模式。當訓練循環反覆依賴 AI 生成嘅內容，而唔係真實世界數據時，呢啲模式會慢慢消失，最後令模型對現實嘅表示出現扭曲。
pmc.ncbi.nlm.nih.govAI models collapse when trained on recursively generated data - PMC
pure.ed.ac.uk[PDF] AI models collapse when trained on recursively generated data
openreview.net[PDF] How Bad is Training on Synthetic Data? A Statistical Analysis of ...
techxplore.comA single real-world data point may stop AI model collapse, analysis suggests