Measuring Moral Inconsistencies in Large Language Models

TL;DR

This paper introduces Semantic Graph Entropy (SGE), a new information-theoretic measure to evaluate the moral consistency of large language models, addressing limitations of previous task-specific accuracy metrics.

Contribution

It proposes SGE as a novel metric for moral consistency in LLMs and demonstrates its better correlation with human judgments compared to existing metrics.

Findings

SGE correlates better with human moral judgments

State-of-the-art LLMs show high moral inconsistency

Enhanced metric using Rules of Thumb improves explanation quality

Abstract

A Large Language Model (LLM) is considered consistent if semantically equivalent prompts produce semantically equivalent responses. Despite recent advancements showcasing the impressive capabilities of LLMs in conversational systems, we show that even state-of-the-art LLMs are highly inconsistent in their generations, questioning their reliability. Prior research has tried to measure this with task-specific accuracy. However, this approach is unsuitable for moral scenarios, such as the trolley problem, with no "correct" answer. To address this issue, we propose a novel information-theoretic measure called Semantic Graph Entropy (SGE) to measure the consistency of an LLM in moral scenarios. We leverage "Rules of Thumb" (RoTs) to explain a model's decision-making strategies and further enhance our metric. Compared to existing consistency metrics, SGE correlates better with human judgments across five LLMs. In the future, we aim to investigate the root causes of LLM inconsistencies and propose improvements.