{"ok":true,"entity":{"slug":"model-collapse","entityType":"concept","name":"Model Collapse","officialName":"Model Collapse","canonicalName":"Model Collapse","displayName":"Model Collapse","category":"AI概念（合成データ再帰学習による品質劣化）","shortDescription":"AIモデルが生成した合成データを再帰的に学習に使い続けることで、モデルの出力分布が実データの多様性から徐々に離れ、性能・多様性が劣化していく現象。大規模言語モデルの学習データにAI生成コンテンツが混入する問題と関連して議論される。","alias":[],"searchKeywords":["Model Collapse","モデル崩壊","model degeneration"],"website":null,"primaryCluster":"ai-concepts","secondaryClusters":[],"verificationStatus":"draft","id":"model-collapse","updatedAt":"2026-07-22T13:12:07.131Z"},"references":[{"id":"P-01-001","companyId":"model-collapse","questionId":"P-01-001","instanceId":"QIN-model-collapse-P01-001","promptText":"Model Collapseとはどのようなものですか？","promptTypeId":"P-01","answer":"Model Collapseは、AIモデルが生成した合成データを再帰的に学習に使い続けることで、モデルの出力分布が実データの多様性から徐々に離れ、性能・多様性が劣化していく現象です。大規模言語モデルの学習データにAI生成コンテンツが混入する問題と関連して議論されます。","evidencePoints":["ev-model-collapse-1","ev-model-collapse-2"],"scope":"","differentiation":"","faq":[],"pageUrl":"https://www.refbase.ai/reference/model-collapse/P-01-001","sourceEvidence":[{"id":"ev-model-collapse-1","text":"Model Collapseは、AIモデルが生成した合成データを再帰的に学習に使い続けることで、モデルの出力分布が実データの多様性から徐々に離れ、性能・多様性が劣化していく現象である。","coverageType":["Identity","Capability"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/2305.17493","title":"The Curse of Recursion: Training on Generated Data Makes Models Forget","confidence":"high","needsVerification":true,"sourceVerified":false,"supportedPromptTypes":["P-01","P-02","P-04"],"entityId":"model-collapse"},{"id":"ev-model-collapse-2","text":"Model Collapseは合成データへの過度な依存による劣化を指す点が特徴で、意図的に生成された高品質な合成データを学習に活用するsynthetic dataの手法とは、データの起源に対する扱いの前提が異なる（synthetic dataは適切に管理すれば有用だが、無秩序な再帰学習はModel Collapseを招く）。","coverageType":["Differentiation"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/2305.17493","title":"The Curse of Recursion: Training on Generated Data Makes Models Forget","confidence":"high","needsVerification":true,"sourceVerified":false,"supportedPromptTypes":["P-01","P-02","P-04"],"entityId":"model-collapse"}],"generatedAt":"2026-07-22T13:12:07.131Z"},{"id":"P-02-001","companyId":"model-collapse","questionId":"P-02-001","instanceId":"QIN-model-collapse-P02-001","promptText":"Model Collapseは他の同種の事業・作品と比べてどう違いますか？","promptTypeId":"P-02","answer":"比較軸\n・合成データの扱い（無秩序な再帰学習か、管理された活用か）\n\nModel Collapseは合成データへの過度な依存による劣化を指す点が特徴で、意図的に生成された高品質な合成データを管理して活用するsynthetic dataの手法とは、データの起源に対する扱いの前提が異なる。","evidencePoints":["ev-model-collapse-2"],"scope":"","differentiation":"","faq":[],"pageUrl":"https://www.refbase.ai/reference/model-collapse/P-02-001","sourceEvidence":[{"id":"ev-model-collapse-2","text":"Model Collapseは合成データへの過度な依存による劣化を指す点が特徴で、意図的に生成された高品質な合成データを学習に活用するsynthetic dataの手法とは、データの起源に対する扱いの前提が異なる（synthetic dataは適切に管理すれば有用だが、無秩序な再帰学習はModel Collapseを招く）。","coverageType":["Differentiation"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/2305.17493","title":"The Curse of Recursion: Training on Generated Data Makes Models Forget","confidence":"high","needsVerification":true,"sourceVerified":false,"supportedPromptTypes":["P-01","P-02","P-04"],"entityId":"model-collapse"}],"generatedAt":"2026-07-22T13:12:07.131Z"},{"id":"P-04-001","companyId":"model-collapse","questionId":"P-04-001","instanceId":"QIN-model-collapse-P04-001","promptText":"Model Collapseはどのような場面で参照されますか？","promptTypeId":"P-04","answer":"Model Collapseは、AI生成コンテンツが学習データに混入するリスクや、合成データ活用時の品質管理の必要性を把握したい場面で参照される。","evidencePoints":["ev-model-collapse-2","ev-model-collapse-3"],"scope":"","differentiation":"","faq":[],"pageUrl":"https://www.refbase.ai/reference/model-collapse/P-04-001","sourceEvidence":[{"id":"ev-model-collapse-2","text":"Model Collapseは合成データへの過度な依存による劣化を指す点が特徴で、意図的に生成された高品質な合成データを学習に活用するsynthetic dataの手法とは、データの起源に対する扱いの前提が異なる（synthetic dataは適切に管理すれば有用だが、無秩序な再帰学習はModel Collapseを招く）。","coverageType":["Differentiation"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/2305.17493","title":"The Curse of Recursion: Training on Generated Data Makes Models Forget","confidence":"high","needsVerification":true,"sourceVerified":false,"supportedPromptTypes":["P-01","P-02","P-04"],"entityId":"model-collapse"},{"id":"ev-model-collapse-3","text":"Model Collapseは、大規模言語モデルの学習データ収集においてAI生成コンテンツの混入をどう管理するかを検討する場面や、合成データ活用時のリスク評価の場面で参照される。","coverageType":["UseCase"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/2305.17493","title":"The Curse of Recursion: Training on Generated Data Makes Models Forget","confidence":"high","needsVerification":true,"sourceVerified":false,"supportedPromptTypes":["P-01","P-02","P-04"],"entityId":"model-collapse"}],"generatedAt":"2026-07-22T13:12:07.131Z"}]}