EulerESG: Automating ESG Disclosure Analysis with LLMs

ESG reports are often published as long, heterogeneous PDF documents, making systematic analysis difficult and labor-intensive. EulerESG presents an LLM-powered system that automates ESG disclosure analysis by explicitly modeling underlying reporting standards, enabling high-fidelity metric extraction and interactive exploration of corporate sustainability data.

Contributions

The paper details several key contributions aimed at addressing the challenges of analyzing unstructured ESG reports:

1. ESG-Standard-Specific Metrics Extraction: The system designs an automated pipeline leveraging LLMs to extract structured ESG metrics from unstructured, noisy, and highly variable corporate reports in PDF format, significantly reducing the need for manual rule engineering.

2. Multi-Framework Standard Alignment: The work covers over 100 industries and multiple reporting frameworks. A specific module enables fine-grained alignment of extracted disclosures with industry-specific requirements, supporting flexible comparisons across companies, industries, and regulatory standards.

3. LLM-Powered Automated ESG Analysis: An ESG agent that integrates LLM reasoning with prompt engineering is designed to perform context-aware entity recognition, metric classification, and compliance flagging. This agent supports benchmarking and decision support through interactive interfaces.

System Design

The EulerESG architecture is decomposed into five specialized modules to manage the complexity of ESG analysis:

1. Standards-Based Metric Extraction & Expansion: This module preprocesses metrics through a three-stage refinement process: first, the Standard Metric Identification (SMI) component extracts relevant metrics from an extensive collection of unstructured ESG standards; second, Standard-Aligned Metrics Refinement augments each metric with metadata; and finally, Semantic Expansion of Metric Definitions enriches each metric with contextual information necessary for vector-based retrieval.

2. Disclosure-Oriented Report Encoding: To handle heterogeneous formats, this module uses a three-stage approach: Content Extraction is performed page by page; followed by Segment-Level Structuring annotates the extracted content at the segment level; and finally, Report Content Embedding transforms each textual segment into a vector representation using an embedding model, specifically adopting the open-source BGE-M3 model.

3. Dual-Channel Metrics Retrieval: This module combines two retrieval methods to find relevant information from reports: Keyword retrieval searches directly for occurrences of metric terminology within the report, and a complementary Semantic Retrieval mechanism, which uses the embeddings generated during the semantic expansion phase to identify relevant segments in the report. Relevance scores are refined by applying vector similarity matching between extended metric definitions and report content using BGE-M3 embeddings, followed by re-ranking with the BGE-Reranker-v2-M3 model.

4. LLM-Driven Disclosure Analysis: This module performs a twophase reasoning process. First, LLM-Guided Content Reasoning employs designed analytical prompts to examine retrieved content segments, evaluating both the indicators and the quality of disclosures based on structured criteria. Second, results undergo Structured MultiClass Disclosure Classification, categorizing each result into one of three predefined classes: (1) (Fully) Discussed/Disclosed, (2) Partially Discussed/Disclosed, or (3) Not Discussed/Disclosed.

5. Interactive LLM ESG Chatbot: This integrated chatbot allows users to interact with reports through natural language queries, supporting on-demand definitions of ESG terms and metric codes, contextual queries about metrics or policies not explicitly covered by standard frameworks, and the summarization of policy narratives or ambiguous disclosures.

Performance Analysis

The system was evaluated on a cross-industry case study using four large companies (BMW, Macquarie Group, P&G, and Dell) across twelve company-industry pairs aligned with SASB standards. The evaluation measured two primary dimensions: metric-level accuracy and running time. Across all tested LLM backends (Gemini-2.5-flash, Deepseek-V3.2-Exp, Claude-Haiku-4.5, GPT-5, and Qwen-3), all models achieved high metric accuracy (Table 1), with the highest average accuracy attained by GPT-5 at 0.95 across the tested pairs. The system demonstrated an efficiency-accuracy trade-off, where while GPT-5 achieved the highest accuracy, models like Gemini-2.5-flash and Qwen-3 offer a balanced middle ground, combining near-top accuracy with substantially lower latency than GPT-5.

Demonstration and Utility

The EulerESG interface is designed as an interactive dashboard featuring a Disclosure Summary Panel, a Metrics Table presenting extracted data in a standardized format, and a PDF Viewer for cross-referencing. The system is positioned to serve diverse stakeholders: Regulators and Investors can use it to monitor ESG compliance, benchmark firms across sectors, and build consistent ESG scorecards. Corporate teams can apply it for pre-publication self-assessment and identifying "missing or weakly substantiated metrics.

Here are specific improvements for existing AI systems based on the EulerESG framework, along with what the improved system could achieve:

1. The current limitation is brittle, rule-based extraction or treating ESG reports as generic text without standard awareness.

2. Improvement: Implement a modular pipeline that explicitly models reporting standards (e.g., SASB, GRI) from the outset using a "Standards-Based Metric Extraction & Expansion" module.

3. What the improved system can do: It will move beyond simple keyword matching to perform high-fidelity extraction of structured metrics, even when the report uses non-standard terminology or different units, by mapping extracted data directly against predefined standard taxonomies.

4. Improvement: Integrate a Dual-Channel Metrics Retrieval mechanism that combines direct Keyword Retrieval with advanced Semantic Retrieval (using embeddings from standardized metric definitions).

5. What the improved system can do: It will significantly reduce false negatives and positives during information retrieval by ensuring that semantic similarity between a reported narrative and a formal metric definition is captured, leading to more comprehensive data capture than keyword search alone.

6. Improvement: Introduce an LLM-Driven Disclosure Analysis module that uses structured analytical prompts to classify disclosures into three granular classes: (1) Fully Discussed/Disclosed, (2) Partially Discussed/Disclosed, and (3) Not Discussed/Disclosed, distinguishing between quantitative disclosure and qualitative discussion.

7. What the improved system can do: It will provide a nuanced assessment of disclosure quality that goes beyond simple presence/absence flags, allowing users to immediately gauge the depth of commitment or risk mitigation described in the report.

8. Improvement: Develop an interactive LLM ESG Chatbot integrated with vector databases representing both ESG standards and extracted report content.

9. What the improved system can do: It will enable natural language querying for complex, contextual insights—such as Compare the governance discussion on board diversity between Company X and Company Y, or Summarize all climate resilience strategies mentioned in Q3 disclosures—providing immediate, context-aware benchmarking and explanation without needing to parse raw documents.

10. Improvement: Implement a dynamic LLM backend selection strategy based on required trade-offs (Efficiency vs. Accuracy).

11. What the improved system can do: It will allow users to select the optimal model for their task—e.g., selecting a smaller, faster model (like Gemini-2.5-flash) for high-throughput screening, or a larger, more capable model (like GPT-5) for deep investigative analysis—optimizing both runtime and required accuracy simultaneously.