DIAL-G²: Graph-Guided Dialectical Agent for Advanced ESG Reasoning

Zechuan Chen, Yanwen Chen, TianMing Sha, Bin Xu, Yongsen Zheng, Keze Wang

Keywords: ESG; Multi-Agent System (MAS)

TL;DR: This paper proposes a novel multi-agent framework where specialized AI agents debate the contents of ESG reports, referencing a shared knowledge base of official standards and company data to generate a final rating.

Abstract: The rapid growth of the focus on Environmental, Social, and Governance (ESG) creates the need for an effective AI tool for evaluating corporate sustainability. Despite the increasing availability of ESG data disclosed by companies, the complex, relational, and unstructured nature of ESG reports poses significant challenges for large-scale LLMs. The surprising efficacy of small language models (SLMs) on domain-specific tasks challenges the prevailing belief that only massive, general-purpose models can tackle complex reasoning. Inspired by this observation, we investigate the potential of compact models in the nuanced field of ESG analysis, arguing that the path to true expertise lies in focused knowledge and structured collaboration. Our work presents a two-fold contribution to realize this vision: 1) We introduce ESGEXPERT-30K, a new knowledge-intensive dataset, to fine-tune a compact SLM into a specialized ESGEEK model that achieves state-of-the-art accuracy on domain-specific QA. 2) We propose DIAL-G², a novel framework where these expert agents form a committee to analyze full, multimodal corporate reports. In DIAL-G², agents populate a shared graph with conflicting arguments. The GNN then performs relational inference on this graph and uses its learned attention weights to direct the agents’ focus towards the most salient and contested information in a subsequent debate phase. This graph-guided approach demonstrates remarkable effectiveness. Furthermore, on our newly contributed, large-scale ESGREPORT-RATING-50K benchmark, DIAL-G² achieves human-expert-level performance in end-to-end ESG score prediction. Extensive experiments show that DIAL-G² can overcome bottlenecks of Multi-Agent Systems (MAS) in domain-specific understanding, providing the ESG research community with a powerful new paradigm for scalable, relational, and interpretable AI.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 5988