ESGenius: Benchmarking LLMs on Environmental, Social, and Governance (ESG) and Sustainability Knowledge

LLMs have significant gaps in ESG and sustainability knowledge: Even state-of-the-art large language models only achieved 55-70% accuracy on ESG questions in zero-shot settings, revealing that these specialized domains remain challenging areas where LLMs cannot be reliably trusted without additional support.

Retrieval-Augmented Generation substantially improves accuracy: When LLMs were provided access to authoritative ESG documents and frameworks through RAG methods, performance improved dramatically—for example, one 14-billion parameter model improved from 63.82% to 80.46% accuracy—demonstrating that grounding AI responses in verified sources is critical for responsible ESG applications.

Smaller models can perform competitively with proper support: The benchmark tested models ranging from 0.5 billion to 671 billion parameters, showing that even smaller models achieved strong performance with RAG, suggesting that responsible ESG AI doesn't necessarily require the largest models if they're properly equipped with reliable information sources.

A comprehensive benchmark now exists for accountability: ESGenius provides 1,136 expert-validated questions linked to 231 authoritative ESG frameworks and standards, creating a transparent evaluation tool that organizations can use to assess whether LLMs are trustworthy for ESG-related decision-making and reporting.

Transparency and source verification are essential: The benchmark's systematic linking of each question to source documents enables organizations to verify where AI-generated ESG information comes from, supporting more responsible deployment of LLMs in sustainability and governance contexts.