Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Advancing ESG Intelligence: An Expert-level Agent and Comprehensive Benchmark for Sustainable Finance

Yilei Zhao, Wentao Zhang, Xiao Lei, Yandan Zheng, Mengpu Liu, Wei Yang Bryan Lim

TL;DR

The paper tackles fragmentation in ESG data and the high-stakes nature of sustainability auditing by introducing ESGAgent, a hierarchical, tool-augmented agent designed to perform professional-grade ESG analysis. It pairs ESGAgent with a comprehensive three-level benchmark derived from 310 DJIA sustainability reports to evaluate both atomic and integrated, standards-aligned reporting tasks. Empirical results show ESGAgent outperforms state-of-the-art closed-source LLM baselines on Level-1/2 tasks and demonstrates depth, verifiability, and structured reporting on Level-3 tasks through a multi-judge evaluation framework. The work provides a rigorous, domain-grounded testbed for advancing agentic capabilities in high-stakes vertical domains and offers transferable insights for other regulated fields requiring data fidelity and reporting rigor.

Abstract

Environmental, social, and governance (ESG) criteria are essential for evaluating corporate sustainability and ethical performance. However, professional ESG analysis is hindered by data fragmentation across unstructured sources, and existing large language models (LLMs) often struggle with the complex, multi-step workflows required for rigorous auditing. To address these limitations, we introduce ESGAgent, a hierarchical multi-agent system empowered by a specialized toolset, including retrieval augmentation, web search and domain-specific functions, to generate in-depth ESG analysis. Complementing this agentic system, we present a comprehensive three-level benchmark derived from 310 corporate sustainability reports, designed to evaluate capabilities ranging from atomic common-sense questions to the generation of integrated, in-depth analysis. Empirical evaluations demonstrate that ESGAgent outperforms state-of-the-art closed-source LLMs with an average accuracy of 84.15% on atomic question-answering tasks, and excels in professional report generation by integrating rich charts and verifiable references. These findings confirm the diagnostic value of our benchmark, establishing it as a vital testbed for assessing general and advanced agentic capabilities in high-stakes vertical domains.

Advancing ESG Intelligence: An Expert-level Agent and Comprehensive Benchmark for Sustainable Finance

TL;DR

The paper tackles fragmentation in ESG data and the high-stakes nature of sustainability auditing by introducing ESGAgent, a hierarchical, tool-augmented agent designed to perform professional-grade ESG analysis. It pairs ESGAgent with a comprehensive three-level benchmark derived from 310 DJIA sustainability reports to evaluate both atomic and integrated, standards-aligned reporting tasks. Empirical results show ESGAgent outperforms state-of-the-art closed-source LLM baselines on Level-1/2 tasks and demonstrates depth, verifiability, and structured reporting on Level-3 tasks through a multi-judge evaluation framework. The work provides a rigorous, domain-grounded testbed for advancing agentic capabilities in high-stakes vertical domains and offers transferable insights for other regulated fields requiring data fidelity and reporting rigor.

Abstract

Environmental, social, and governance (ESG) criteria are essential for evaluating corporate sustainability and ethical performance. However, professional ESG analysis is hindered by data fragmentation across unstructured sources, and existing large language models (LLMs) often struggle with the complex, multi-step workflows required for rigorous auditing. To address these limitations, we introduce ESGAgent, a hierarchical multi-agent system empowered by a specialized toolset, including retrieval augmentation, web search and domain-specific functions, to generate in-depth ESG analysis. Complementing this agentic system, we present a comprehensive three-level benchmark derived from 310 corporate sustainability reports, designed to evaluate capabilities ranging from atomic common-sense questions to the generation of integrated, in-depth analysis. Empirical evaluations demonstrate that ESGAgent outperforms state-of-the-art closed-source LLMs with an average accuracy of 84.15% on atomic question-answering tasks, and excels in professional report generation by integrating rich charts and verifiable references. These findings confirm the diagnostic value of our benchmark, establishing it as a vital testbed for assessing general and advanced agentic capabilities in high-stakes vertical domains.
Paper Structure (42 sections, 5 figures, 6 tables)

This paper contains 42 sections, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. ESG Analytics and Methods
  4. Benchmarking Agentic AI Systems
  5. ESGAgent
  6. Why this agent?
  7. Architecture of Agent
  8. ESG-Oriented Toolset
  9. ESG Benchmark
  10. Why this benchmark?
  11. Composition of Benchmark
  12. Datasets
  13. Capabilities
  14. Experiment
  15. Evaluation
  16. ...and 27 more sections

Figures (5)

  • Figure 1: The synergistic interaction between ESG Agent and Benchmark.
  • Figure 2: ESGAgent overall architecture and workflow.
  • Figure 3: Samples of ESG questions. The benchmark progresses from static questions (Level 1 & 2) to an open-ended, template-driven architecture (Level 3).
  • Figure 4: Distribution of ESG labels and required agent capabilities.
  • Figure B.5: (Top-left, Bottom-left, Bottom-right) Ranking of different models on level 1 and 2 tasks. (Top-right) Evaluation heatmap on level 1 and 2 tasks (18 models $\times$ 246 questions).