The Accuracy, Robustness, and Readability of LLM-Generated Sustainability-Related Word Definitions

Heiman, Alice

arXiv.org Artificial Intelligence 

Thus, this can lead to inconsistencies in research and policy-making. A common language with standardized To address this issue, the Interdisciplinary Panel definitions is crucial for effective climate on Climate Change (IPCC) and the United Nations discussions. However, concerns exist (UN) maintain the online glossaries IPCC about LLMs misrepresenting climate Glossary (IPCC, 2019a,b, 2018), and UNTERM terms. We compared 300 official IPCC (UN, 2024a). Although LLMs have access to glossary definitions with those generated these repositories during training, they are not by GPT-4o-mini, Llama3.1 8B, and Mistral constrained to them during inference. Therefore, 7B, analyzing adherence, robustness, LLMs could further diversify and confuse these and readability using SBERT sentence embeddings.