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

Yilei Zhao, Wentao Zhang, Lei Xiao, Yandan Zheng, Mengpu Liu, Chenyue Song

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.

Introduction

Professional Environmental, Social, and Governance (ESG) analysis is a critical yet challenging domain within sustainable finance. Analysts must navigate a landscape of fragmented data buried in thousands of pages of unstructured PDFs, financial filings, and multimedia sources. Furthermore, the workflow requires complex multi-step reasoning—such as reconciling divergent ratings or calculating portfolio-wide carbon intensity—that standard large language models (LLMs) often fail to execute reliably.

To bridge this gap, we introduce ESGAgent, a hierarchical multi-agent system designed to automate complex ESG auditing. Alongside this system, we present a rigorous three-level benchmark derived from over a decade of corporate disclosures, challenging AI to move beyond simple Q&A toward professional-grade report generation.

The ESGAgent Framework

The core of our approach is a hierarchical multi-agent architecture that functions like a professional research team. Rather than relying on simple prompting, the system empowers a suite of specialized tools and sub-agents to handle specific aspects of the ESG audit workflow.

Figure 3: Detailed architecture of the ESGAgent, illustrating the workflow from query processing to final report generation using specialized analytical tools.

Key components of the architecture include:

  • Central Planner: Decomposes high-level user queries into discrete, executable sub-tasks.
  • Specialized Toolset: Utilizes LightRAG, a retrieval-augmented generation framework that combines Vector Databases with Knowledge Graphs to ensure longitudinal consistency across years of data.
  • Domain-Specific Sub-Agents: Includes a Deep Researcher for web searching, a Python Interpreter for financial and carbon calculations, a Plotter for chart generation, and a Deep Analyzer for multi-source synthesis.
  • Iterative Refinement: The system employs a re-execution loop to refine outputs if a sub-agent fails to meet the initial task target.

A Comprehensive ESG Benchmark

To evaluate agentic performance in this high-stakes vertical, we developed a benchmark based on 310 official reports from Dow Jones Industrial Average (DJIA) constituents spanning 2010–2024. The benchmark is structured into three distinct difficulty tiers to test different cognitive and technical capabilities.

Figure 6: The three-level benchmark structure, ranging from atomic literacy (Level 1) to multimodal reasoning (Level 2) and professional report generation (Level 3).
  • Level 1 (Easy): 132 atomic tasks testing fundamental ESG literacy (e.g., CFA-style questions).
  • Level 2 (Medium): 114 compositional tasks requiring 3–5 reasoning steps and multimodal processing of PDFs, CSVs, and audio/video files.
  • Level 3 (Advanced): 45 open-ended tasks requiring the generation of professional-grade ESG reports aligned with global standards like GRI or SASB.

Key Insights and Results

Empirical evaluation reveals that ESGAgent significantly outperforms existing baselines. It achieved an average accuracy of 84.15% on Level 1 and 2 tasks, surpassing Gemini-3-flash (80.89%). In Level 3 evaluations, judged by an ensemble of frontier LLMs, ESGAgent scored 8.096/10 on average, excelling in factual consistency and presentation quality.

Our analysis highlights a "Cognitive Shift": standard models show clear performance decay when moving from simple fact-retrieval (Level 1) to tool-based reasoning (Level 2). Furthermore, ESGAgent prioritizes information density over verbosity, producing reports rich in verifiable references and visualizations rather than "hallucinated fluff."

Limitations and Future Work

While highly effective, ESGAgent currently faces operational latency, taking approximately 10 minutes to generate a high-quality Level 3 report. Future work will focus on reducing computational costs and expanding the benchmark to include international regulatory frameworks, such as the EU's Corporate Sustainability Reporting Directive (CSRD). Additionally, incorporating human-in-the-loop validation from professional ESG auditors will further refine the system's diagnostic accuracy.

Summary updated

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.
BibTeX
@misc{zhao2026advancing,
  author = {Yilei Zhao and Wentao Zhang and Lei Xiao and Yandan Zheng and Mengpu Liu and Chenyue Song},
  title = {{Advancing ESG Intelligence: An Expert-level Agent and Comprehensive Benchmark for Sustainable Finance}},
  year = {2026},
  eprint = {2601.08676},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.08676},
}
Figure