{ "format": "agent-article-v0", "article_id": "EA-122", "canonical_url": "/tools/ea-122-mixture-of-agents", "page_type": "reference", "claim_ids": [], "evidence_ids": [ "EVD-EA-122-URL-99D984CF631B", "EVD-EA-122-URL-16E11E26BE5B", "EA-122-R11-RESEARCH", "EA-122-R11-WEB-READBACK" ], "unknown_ids": [ "EA-122-UNKNOWN-01", "EA-122-UNKNOWN-CURRENTNESS" ], "operational_object_ids": [], "provider": [], "models": [], "plans": [], "surfaces": [], "regions": [], "current_status": "UNKNOWN", "observation_status": "SPEC_ONLY", "freshness": "UNKNOWN", "last_verified": null, "supersedes": [], "superseded_by": [], "provenance": [ "COMM-042-EA0122", "SRC-EA-122-URL-99D984CF631B", "SRC-EA-122-URL-16E11E26BE5B", "WEBBODY-EA0122-R8-F149E374" ], "article_sources": [ { "id": "EVD-EA-122-URL-99D984CF631B", "evidence_class": "RESEARCH", "title": "source class: primary research paper; exact title: Mixture-of-Agents Enhances Large Language Model Capabilities; author/org: Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, Jame", "href": "https://arxiv.org/abs/2406.04692", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-122-URL-99D984CF631B" ], "supporting_locators": [ "EVIDENCE TABLE", "- source class: primary research paper; exact title: Mixture-of-Agents Enhances Large Language Model Capabilities; author/org: Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, James Zou; canonical URL: https://arxiv.org/abs/2406.04692" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EVD-EA-122-URL-16E11E26BE5B", "evidence_class": "RESEARCH", "title": "source class: primary research paper; exact title: Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena; author/org: Lianmin Zheng et al.; canonical URL:", "href": "https://arxiv.org/abs/2306.05685", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-122-URL-16E11E26BE5B" ], "supporting_locators": [ "EVIDENCE TABLE", "- source class: primary research paper; exact title: Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena; author/org: Lianmin Zheng et al.; canonical URL: https://arxiv.org/abs/2306.05685" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EA-122-R11-RESEARCH", "evidence_class": "RESEARCH", "title": "EA-122 監査済みresearch / canonical revision binding", "supports": [ "MoA原著は、複数LLMをlayerに配置し、前layerの全出力を次layerが補助情報として受け取る構造を明示し、特定benchmarkでの結果を報告する。したがってparallel generation→aggregationの実装パターンと、quality/cost評価の必要性は確立する。しかし、報告scoreは原著のモデル組合せ・prompt・dataset・judge条件に依存し、業務正確性、費用対効果、低遅延、全providerでの優越を確立しない。judge論文が示す評価biasも併記すべきである。" ], "limit": "canonical監査・revision bindingの存在を示す。外部ソースcurrentnessが本文またはcontrol-planeで確定していない場合は、currentnessをUNKNOWNのまま保持する。", "origin_class": "AUDITED_CANONICAL_RESEARCH_PACKET", "source_or_experiment_ids": [ "COMM-042-EA0122" ], "supporting_locators": [ "COMM-042-EA0122", "EA-122-R5-LEGACY-MIGRATION-20261008-R1", "content/staging/EA-122.md" ], "scope": "監査済みEA本文・canonical revision・既存Source Graph relationの範囲。欠損currentnessは推測で閉じない。", "provenance_status": "AUDIT_PASS_CANONICAL_BINDING" }, { "id": "EA-122-R11-WEB-READBACK", "evidence_class": "IMPL", "title": "EA-122 Web body identity / canonical readback", "supports": [ "canonical staging bodyのSHA-256一致とR9 canonical ledger readbackをProduction projectionの入力として検証する。" ], "limit": "body identity/readbackは本文claim自体の真偽や外部ソースの最新性を追加保証しない。", "checked_at": null, "origin_class": "IMPLEMENTATION_READBACK", "source_or_experiment_ids": [ "WEBBODY-EA0122-R8-F149E374" ], "supporting_locators": [ "WEBBODY-EA0122-R8-F149E374", "content/staging/EA-122.md", "EA-122-R5-LEGACY-MIGRATION-20261008-R1" ], "scope": "R9/R10 canonical readback後のProduction生成時body identity。", "provenance_status": "CANONICAL_READBACK_BODY_HASH_MATCH" } ], "research_sections": [ { "heading": "WHY_THIS_MATTERS_TO_UPDATE", "text": "並列LLMの出力を別agentが統合するMoAは、品質向上と費用/遅延のトレードオフを明示して初めて業務設計になる。benchmarkの勝利を企業保証へ外挿しない。" }, { "heading": "SCOPE_ROUTING", "text": "multi-LLM / evaluation / implementation / cost / interoperability" }, { "heading": "CLAIM_OR_RESEARCH_QUESTION", "text": "MoA原著が示すlayered architectureと評価結果は何か。複数出力の統合が常に単一modelを上回るか。" }, { "heading": "EVIDENCE TABLE", "text": "- source class: primary research paper; exact title: Mixture-of-Agents Enhances Large Language Model Capabilities; author/org: Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, James Zou; canonical URL: https://arxiv.org/abs/2406.04692; publication/update/access: submitted 2024-06-07 / accessed 2026-09-13 JST; version/release/tag/commit/API revision/model ID: arXiv v1; supporting locator: abstract lines 8-20（layered MoA、各layerの複数LLM agent、前layer全出力をauxiliary informationとして利用、AlpacaEval 2.0/MT-Bench/FLASK、reported 65.1% vs 57.5% on AlpacaEval 2.0）; evidence label: RESEARCH; relation: support / limit.\n- source class: primary research paper; exact title: Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena; author/org: Lianmin Zheng et al.; canonical URL: https://arxiv.org/abs/2306.05685; publication/update/access: submitted 2023-06-09, revised v4 2023-12-24 / accessed 2026-09-13 JST; version/release/tag/commit/API revision/model ID: arXiv v4 / NeurIPS 2023 Datasets and Benchmarks Track; supporting locator: abstract lines 16-23（MT-Bench/Chatbot Arena、judge biases、human preference agreement、benchmark scope）; evidence label: RESEARCH; relation: limit / context." }, { "heading": "FINDINGS", "text": "MoA原著は、複数LLMをlayerに配置し、前layerの全出力を次layerが補助情報として受け取る構造を明示し、特定benchmarkでの結果を報告する。したがってparallel generation→aggregationの実装パターンと、quality/cost評価の必要性は確立する。しかし、報告scoreは原著のモデル組合せ・prompt・dataset・judge条件に依存し、業務正確性、費用対効果、低遅延、全providerでの優越を確立しない。judge論文が示す評価biasも併記すべきである。" }, { "heading": "LIMITS / COUNTEREVIDENCE / FAILURE ENVELOPE", "text": "すべての候補が同じ誤りを共有すると統合は誤りを強化する。layer追加はtoken、latency、failure surface、judge/aggregator dependencyを増やす。AlpacaEval/MT-Bench/FLASKの勝利は、メール・契約・RAG・codeの安全性を示さない。" }, { "heading": "AVAILABILITY_GATES", "text": "各model/API/license/region/quota、parallelism、aggregator model、context window、provider data policy、budget、traceを確認する。実provider混成run未実施。" }, { "heading": "IMPLEMENTATION_STATE", "text": "bounded fan-out、schema-constrained candidate outputs、diversity check、aggregator、abstain、cost/latency budget、human reviewを設計候補とする。" }, { "heading": "BENCHMARK_OR_EVAL", "text": "single vs MoAで、task success、factuality/citation、candidate diversity、aggregation error、token/tool cost、p50/p95 latency、failure correlation、human preferenceを別測定。judge bias correctionを含める。実測なし。" }, { "heading": "SECURITY / PRIVACY / AUDIT / COST / OPERABILITY", "text": "SECURITY/PRIVACY: candidates間で秘密を複製しない、aggregatorへ最小context。AUDIT: model role/version、layer graph、candidate hashes、judge prompt、cost、latency。COST: fan-out×layers×aggregatorを計上。OPERABILITY: max layer、timeout、partial quorum、fallback single model。" }, { "heading": "SAFE_WORDING_FOR_FUTURE_BOOK", "text": "「MoA原著はlayeredな複数LLMと出力統合を提案し、特定benchmarkでの改善を報告する。ただし、同じ構成が全業務・全providerで優越する証拠ではなく、品質・相関誤り・遅延・token費用・judge biasを含む比較が必要である。」" }, { "heading": "UPDATE_WORK_HANDOFF", "text": "本文ではMoAのtopology、aggregation rule、benchmark、費用、可用性を別欄にし、原著のbenchmark値にはdataset/model/judge/versionを付す。現行MoA実装はOSS release/commitへ追跡する。" }, { "heading": "OPEN_UNKNOWN", "text": "業務別MoAのcost-quality frontier、diversityの測定、aggregator誤り、judge-independent評価、provider/API更新、privacy複製範囲。" }, { "heading": "STOP_CONDITION", "text": "reported AlpacaEval/MT-Bench/FLASK scoreを本番業務の正確性・安全性・ROIへ変換しない。\nUNIT_CLOSE_STATUS: INDIVIDUAL_BOUNDED_CLOSE / EVIDENCE_RECORDED / UNKNOWN_PRESERVED\n\n---" } ], "human_url": "/tools/ea-122-mixture-of-agents", "markdown_url": "/agent/article.md?path=%2Ftools%2Fea-122-mixture-of-agents", "json_url": "/agent/article.json?path=%2Ftools%2Fea-122-mixture-of-agents" }