International Journal of Science, Engineering and TechnologyAuthors: Mary Shakkina C, Miranda Lakshmi Travis, Martin Aruldoss\
Abstract: Large language models are highly capable at generating text, reasoning, answering questions, summarizing content, and supporting decisions. However, understanding how these systems arrive at their outputs remains difficult for anyone other than specialist researchers. Prior work on making artificial intelligence understandable has largely focused on producing explanations that are easy for people to follow, such as showing where the model pays attention or highlighting which parts of the input appear important. An explanation that is easy to understand, convincing, or simple is not necessarily faithful to how the model actually arrived at its answer. This is a significant problem for high-stakes applications of large language models, where people may come to trust an incorrect answer more because it is accompanied by an explanation that sounds convincing but is not actually true.This paper proposes the Causal, Layered, Audience-aware Explanation framework (CLAX) for evaluating explanations of large language models. We identify four considerations for evaluating explanations that previous research has treated together rather than separately: how easy the explanation is to understand, how well it matches the model's behaviour, how faithful it is to the model's actual reasoning, and how useful it is in practice. We identify five levels that explanations can target, ranging from which inputs are important to how the model encodes, generates, and behaves in the world. We present a set of criteria for evaluating explanations, including assessing the impact of different inputs, ensuring that the explanation is complete and correct, checking that it is consistent under variations in the input, determining whether it is understandable to people, and explicitly stating what remains unknown. Rather than attempting to capture explanation quality in a single metric, CLAX highlights the importance of considering the audience, content, evidence, utility, and limitations of explanations together. CLAX is applicable to any language model, whether or not its internal mechanisms are observable, which makes it possible to reproduce research on making language models understandable. We also propose a way to evaluate this framework in future studies across different models, research areas, and types of explanations in order to determine how well it performs.
DOI: https://doi.org/10.5281/zenodo.22202179
