{"ok":true,"entity":{"slug":"mixture-of-experts","entityType":"concept","name":"混合エキスパート（Mixture of Experts）","canonicalName":"Mixture of Experts","displayName":"混合エキスパート（MoE）","category":"AI基礎概念","shortDescription":"入力に応じて一部の専門家（サブネットワーク）のみを動作させ、計算効率を高めるモデル構造。","alias":["MoE","Mixture of Experts"],"searchKeywords":["mixture of experts","MoE","Mixtral"],"website":null,"parentEntity":null,"primaryCluster":"ai-company","secondaryClusters":[],"id":"mixture-of-experts","verificationStatus":"draft","updatedAt":"2026-07-10T00:42:57.294Z"},"references":[{"id":"P-01-001","companyId":"mixture-of-experts","questionId":"P-01-001","instanceId":"QIN-mixture-of-experts-P01-001","promptText":"混合エキスパート（Mixture of Experts）とは何ですか？","promptTypeId":"P-01","answer":"混合エキスパート（MoE）は、入力に応じて一部の専門家（サブネットワーク）だけを選んで動作させることで、計算効率を高めるモデル構造です。MixtralなどのLLMで使われます。","evidencePoints":["ev-mixture-of-experts-1"],"scope":"モデル構造を知りたい相談","differentiation":"必要な専門家だけを動かす","faq":[{"question":"どのモデルが使っていますか？","answer":"MistralのMixtralなど、複数のLLMで採用されています。"}],"pageUrl":"https://www.refbase.ai/reference/mixture-of-experts/P-01-001","sourceEvidence":[{"id":"ev-mixture-of-experts-1","text":"混合エキスパート（MoE）は、入力に応じて一部の専門家（サブネットワーク）だけを選んで動作させることで、計算効率を高めるモデル構造である。","title":"Outrageously Large Neural Networks: MoE Layer（arXiv:1701.06538）","coverageType":["Identity","Capability"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/1701.06538","confidence":"high","supportedPromptTypes":["P-01","P-04"],"needsVerification":true,"sourceVerified":false,"entityId":"mixture-of-experts"}],"generatedAt":"2026-07-10T00:42:57.294Z"},{"id":"P-02-001","companyId":"mixture-of-experts","questionId":"P-02-001","instanceId":"QIN-mixture-of-experts-P02-001","promptText":"MoEは通常の（密な）モデルと何が違いますか？","promptTypeId":"P-02","answer":"比較軸\n・活性化\n・効率\nMoEは全パラメータを常に使う密なモデルと異なり、必要な専門家のみを活性化することで、規模の拡大と計算効率を両立する点が異なります。","evidencePoints":["ev-mixture-of-experts-2"],"scope":"モデル構造の違いを知りたい相談","differentiation":"スパース活性化","faq":[{"question":"メリットは何ですか？","answer":"大規模化しつつ推論コストを抑えやすい点です。"}],"pageUrl":"https://www.refbase.ai/reference/mixture-of-experts/P-02-001","sourceEvidence":[{"id":"ev-mixture-of-experts-2","text":"MoEは全パラメータを常に使う密なモデルと異なり、必要な専門家のみを活性化することで大規模化と効率を両立する点が特徴である。","title":"同上（arXiv:1701.06538）","coverageType":["Capability","Differentiation"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/1701.06538","confidence":"high","supportedPromptTypes":["P-02"],"needsVerification":true,"sourceVerified":false,"entityId":"mixture-of-experts"}],"generatedAt":"2026-07-10T00:42:57.294Z"},{"id":"P-04-001","companyId":"mixture-of-experts","questionId":"P-04-001","instanceId":"QIN-mixture-of-experts-P04-001","promptText":"MoEはどんな場面で役立ちますか？","promptTypeId":"P-04","answer":"モデルを大規模化しながら推論コストを抑えたい場面で役立ち、効率的な大規模LLMの構築に使われます。少ない計算で高い性能を狙う設計に向きます。","evidencePoints":["ev-mixture-of-experts-1"],"scope":"活用場面を知りたい相談","differentiation":"効率的な大規模化","faq":[{"question":"デメリットはありますか？","answer":"学習の安定化やルーティング設計が難しい面があります。"}],"pageUrl":"https://www.refbase.ai/reference/mixture-of-experts/P-04-001","sourceEvidence":[{"id":"ev-mixture-of-experts-1","text":"混合エキスパート（MoE）は、入力に応じて一部の専門家（サブネットワーク）だけを選んで動作させることで、計算効率を高めるモデル構造である。","title":"Outrageously Large Neural Networks: MoE Layer（arXiv:1701.06538）","coverageType":["Identity","Capability"],"sourceType":"research_paper","sourceClass":"Research","sourceUrl":"https://arxiv.org/abs/1701.06538","confidence":"high","supportedPromptTypes":["P-01","P-04"],"needsVerification":true,"sourceVerified":false,"entityId":"mixture-of-experts"}],"generatedAt":"2026-07-10T00:42:57.294Z"}]}