ReGMS: Retrieval-Grounded Multi-Agent Scenario Analysis for Climate Risk
ReGMS: 検索に基づくマルチエージェント気候リスクシナリオ分析 (AI 翻訳)
Yun-Kae Kiang, King H. Lam
🤖 gxceed AI 要約
日本語
本論文は、気候シナリオ分析をマルチエージェントシステムとして捉え、IFRS S2/TCFD報告に対応する手法ReGMSを提案。検索型エージェントが公開シナリオを取得し、設計・定量化・コンプライアンス検証の各エージェントが協調してシナリオを構築する。NGFS経路とNASA気候データを用いた実験で、内部一貫性と基準適合性を確認した。
English
This paper proposes ReGMS, a retrieval-grounded multi-agent architecture for climate scenario analysis to support IFRS S2/TCFD reporting. Specialized LLM agents coordinate to build and verify transition and physical-risk scenarios. Using NGFS Phase V pathways and NASA climate data, experiments show internal consistency and standards alignment.
Unofficial AI-generated summary based on the public title and abstract. Not an official translation.
📝 gxceed 編集解説 — Why this matters
日本のGX文脈において
日本でもSSBJ基準の策定が進み、気候シナリオ分析は有報記載や投資家対応で重要性が増している。本手法は、公開データに基づく透明なシナリオ作成を可能にし、日本の企業がTCFD/IFRS S2対応を効率化する上で実用的な指針となる。
In the global GX context
TCFD and ISSB reporting require transparent, traceable scenario analysis. ReGMS offers a novel multi-agent LLM approach that grounds scenarios in public pathways (NGFS) and climate data, enabling firms to meet disclosure standards with verifiable assumptions. This advances the global practice of climate risk assessment.
👥 読者別の含意
🔬研究者:Demonstrates a novel multi-agent architecture for automated scenario analysis, with convergence guarantees and empirical comparison to baselines.
🏢実務担当者:Provides a potential tool to automate parts of TCFD/ISSB scenario analysis reporting, improving traceability and consistency.
🏛政策担当者:Illustrates how AI can support regulatory compliance with scenario analysis requirements, potentially shaping future disclosure guidelines.
📄 Abstract（原文）
We study climate--related scenario analysis through the lens of multi--agent systems. Our goal is to support IFRS S2/TCFD reporting with scenarios that are grounded in public pathways and transparent in their assumptions. We present ReGMS, a retrieval--grounded agent architecture where specialized LLM agents (retrieval, scenario design, quantitative coupling, compliance) coordinate to build and verify transition and physical--risk scenarios. We cast the task as a constrained stochastic game in which retrieved evidence sets feasible actions and verification checks. We discuss convergence behavior of evidence--constrained best--response updates and empirically compare against a centralized planner baseline. For evaluation, we use the Network for Greening the Financial System (NGFS) Phase V transition pathways and NASA NEX--GDDP--CMIP6 downscaled daily projections (from which we derive simple heat and rainfall indices). We report diagnostics for internal consistency against NGFS reference trajectories, scenario diversity across canonical regulatory cases (Current Policies vs.\ Net~Zero~2050), and standards alignment with citation provenance. Baselines include an IAM--style planner, independent agents, and a single--LLM pipeline. These experiments illustrate that ReGMS can produce traceable, standards--aligned scenarios under explicit constraints.
🔗 Provenance — このレコードを発見したソース
- semanticscholar https://doi.org/10.65109/ravt8220first seen 2026-07-19 05:36:08
🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録（無料・3キーワードまで）。
gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。