{
  "format": "agent-article-v0",
  "article_id": "EA-123",
  "canonical_url": "/tools/ea-123-model-debate-deliberation",
  "page_type": "reference",
  "claim_ids": [],
  "evidence_ids": [
    "EVD-EA-123-URL-992190A4815D",
    "EA-123-R11-RESEARCH",
    "EA-123-R11-WEB-READBACK"
  ],
  "unknown_ids": [
    "EA-123-UNKNOWN-01",
    "EA-123-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-EA0123",
    "SRC-EA-123-URL-992190A4815D",
    "WEBBODY-EA0123-R8-789DDB98"
  ],
  "article_sources": [
    {
      "id": "EVD-EA-123-URL-992190A4815D",
      "evidence_class": "RESEARCH",
      "title": "source class: primary research paper; exact title: Improving Factuality and Reasoning in Language Models through Multiagent Debate; author/org: Yilun Du, Shuang Li, Antonio Torralb",
      "href": "https://arxiv.org/abs/2305.14325",
      "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-123-URL-992190A4815D"
      ],
      "supporting_locators": [
        "EVIDENCE TABLE",
        "- source class: primary research paper; exact title: Improving Factuality and Reasoning in Language Models through Multiagent Debate; author/org: Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, Igor Mordatch; canonical URL: https://arxiv.org/abs/2305.14325"
      ],
      "scope": "EVIDENCE TABLE",
      "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED"
    },
    {
      "id": "EA-123-R11-RESEARCH",
      "evidence_class": "RESEARCH",
      "title": "EA-123 監査済みresearch / canonical revision binding",
      "supports": [
        "原著は、複数LLM instanceが個別回答・reasoningを提案し、複数roundでdebateして共通final answerへ到達する手法を提示し、論文内の複数taskで数学/戦略推論とfactual validityの改善を報告する。確立されるのは論文が設定した手続・task・model条件下の研究結果であり、debateが真理判定機関になることではない。合意は独立な証拠・外部検証・人間監督の代替ではない。"
      ],
      "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-EA0123"
      ],
      "supporting_locators": [
        "COMM-042-EA0123",
        "EA-123-R5-LEGACY-MIGRATION-20261008-R1",
        "content/staging/EA-123.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-123-R11-WEB-READBACK",
      "evidence_class": "IMPL",
      "title": "EA-123 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-EA0123-R8-789DDB98"
      ],
      "supporting_locators": [
        "WEBBODY-EA0123-R8-789DDB98",
        "content/staging/EA-123.md",
        "EA-123-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": "複数モデルに提案・反証・再回答させる方式は、単一回答の見落としを検出し得るが、合意形成を正しさと誤認すると誤りの相互強化を招く。"
    },
    {
      "heading": "SCOPE_ROUTING",
      "text": "multi-LLM / evaluation / governance / implementation / cost"
    },
    {
      "heading": "CLAIM_OR_RESEARCH_QUESTION",
      "text": "Multiagent Debate原著は、複数instanceの複数round debateがreasoning/factualityへどの効果を示したか。debate回数・合意・モデル差を一般業務へ適用できるか。"
    },
    {
      "heading": "EVIDENCE TABLE",
      "text": "- source class: primary research paper; exact title: Improving Factuality and Reasoning in Language Models through Multiagent Debate; author/org: Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, Igor Mordatch; canonical URL: https://arxiv.org/abs/2305.14325; publication/update/access: submitted 2023-05-23 / accessed 2026-09-13 JST; version/release/tag/commit/API revision/model ID: arXiv v1; supporting locator: abstract lines 8-24（multiple instances propose/debate over rounds、mathematical/strategic reasoning、factual validity/hallucination claims、black-box models）; evidence label: RESEARCH; relation: support / limit."
    },
    {
      "heading": "FINDINGS",
      "text": "原著は、複数LLM instanceが個別回答・reasoningを提案し、複数roundでdebateして共通final answerへ到達する手法を提示し、論文内の複数taskで数学/戦略推論とfactual validityの改善を報告する。確立されるのは論文が設定した手続・task・model条件下の研究結果であり、debateが真理判定機関になることではない。合意は独立な証拠・外部検証・人間監督の代替ではない。"
    },
    {
      "heading": "LIMITS / COUNTEREVIDENCE / FAILURE ENVELOPE",
      "text": "同じ誤前提・学習偏り・prompt injectionを共有するagent同士は誤りを補強し得る。多数派/強い話者への追従、round増加によるコストと遅延、最終合意によるuncertainty消失は設計上のfailure envelopeであり、本原著abstractから発生率は確定できない（INFERENCE/UNKNOWN）。"
    },
    {
      "heading": "AVAILABILITY_GATES",
      "text": "model/API availability、同時実行、context budget、debate protocol、external evidence/tool access、human escalation、data policyを確認する。実debate run未実施。"
    },
    {
      "heading": "IMPLEMENTATION_STATE",
      "text": "independent proposal→evidence challenge→rebuttal→judge/verification→abstainという段階を設計候補とし、合意しない経路を残す。実装・task-specific calibration未実施。"
    },
    {
      "heading": "BENCHMARK_OR_EVAL",
      "text": "single/majority/debateを同じfixtureで比較し、answer correctness、evidence entailment、counterexample discovery、agreement、abstention、turn/token/latency costを分離。合意率を正確性と同一視しない。"
    },
    {
      "heading": "SECURITY / PRIVACY / AUDIT / COST / OPERABILITY",
      "text": "SECURITY/PRIVACY: candidate間のsecret redaction、untrusted evidence隔離。AUDIT: round graph、model/version、claims、counterclaims、judge、human decision。COST: rounds×agents×judge。OPERABILITY: max rounds、early stop only with evidence, disagreement queue。"
    },
    {
      "heading": "SAFE_WORDING_FOR_FUTURE_BOOK",
      "text": "「Multiagent Debate原著は複数LLM instanceの複数round提案・反証が特定taskの推論/factualityを改善し得ると報告した。debateの合意は外部証拠・独立検証・人間監督の代替ではなく、業務適用には反証・abstain・費用上限が必要である。」"
    },
    {
      "heading": "UPDATE_WORK_HANDOFF",
      "text": "本文改訂では debate topology、独立性、外部evidence、judge、disagreement、停止条件を明示し、原著のtask/model/round条件を付記する。"
    },
    {
      "heading": "OPEN_UNKNOWN",
      "text": "debateの最適round、モデル独立性、誤り相関、domain-specific evidence、long-horizon agentでの効果、費用対効果、企業データ漏えい。"
    },
    {
      "heading": "STOP_CONDITION",
      "text": "consensus・多数決・流暢なrebuttalを正しさの証明としない。論文の改善結果を全業務・全モデルへ一般化しない。\nUNIT_CLOSE_STATUS: INDIVIDUAL_BOUNDED_CLOSE / EVIDENCE_RECORDED / UNKNOWN_PRESERVED\n\n---"
    }
  ],
  "human_url": "/tools/ea-123-model-debate-deliberation",
  "markdown_url": "/agent/article.md?path=%2Ftools%2Fea-123-model-debate-deliberation",
  "json_url": "/agent/article.json?path=%2Ftools%2Fea-123-model-debate-deliberation"
}