Large Language Models Often Say One Thing and Do Another

TL;DR

The experimental results indicate that alignment only on words or deeds poorly and unpredictably influences the other aspect, which supports the hypothesis that the underlying knowledge guiding LLMs' word or deed choices is not contained within a unified space.

Abstract

As large language models (LLMs) increasingly become central to various applications and interact with diverse user populations, ensuring their reliable and consistent performance is becoming more important. This paper explores a critical issue in assessing the reliability of LLMs: the consistency between their words and deeds. To quantitatively explore this consistency, we developed a novel evaluation benchmark called the Words and Deeds Consistency Test (WDCT). The benchmark establishes a strict correspondence between word-based and deed-based questions across different domains, including opinion vs. action, non-ethical value vs. action, ethical value vs. action, and theory vs. application. The evaluation results reveal a widespread inconsistency between words and deeds across different LLMs and domains. Subsequently, we conducted experiments with either word alignment or deed alignment to observe their impact on the other aspect. The experimental results indicate that alignment only on words or deeds poorly and unpredictably influences the other aspect. This supports our hypothesis that the underlying knowledge guiding LLMs' word or deed choices is not contained within a unified space.

Figures (11)

• Figure 1: Illustrations of consistency (left) and inconsistency (right) between LLMs' words and deeds. In this paper, the term "word" specifically refers to the stated opinions, values, or other beliefs of LLMs, while "deed" refers to their actions in specific situations. It is common for LLMs to say one thing and do another.
• Figure 2: The construction pipeline of Deed questions, which involves three main components: the situation, a fixed question and action options. Each element of the Deed questions is generated by GPT-4. Arrows between these elements indicate the flow of input and output within the model.
• Figure 3: The effects of separate word alignment (the first row) or deed alignment (the second row) on another. Two metrics are assessed: direct change rate, the proportion of responses that change following direct alignment and indirect change rate, the proportion of responses that change due to indirect influences, categorized as consistent or inconsistent before alignment. The axes Si and Di represent the ith epoch in SFT and DPO training, respectively.
• Figure 4: The effect of alignment difficulty on generalization.
• Figure 5: The proportion of instances where LLMs maintained a consistent stance across five trials at different temperature settings.