Abstract
Survey-style evaluations of large language models often treat a prompted response as a measure of a model's values or beliefs. This assumption is particularly fragile when responses are read as evidence of political values, social attitudes, or beliefs. We investigate whether prompt robustness differs between objective questions with fixed answers and subjective questions that inquire about opinions or values.
We evaluate four instruction-tuned model families on three objective datasets (MMLU, ARC, and CulturalBench) and three subjective datasets (Political Compass Test, ValueBench, and World Values Survey). For each question/statement, we apply multiple types of prompt changes, such as variations in wording, framing, and format, and measure whether the model provides the same answer across variants.
Using a binomial generalized estimating equation, we find significant effects of model, dataset, prompt category, and their interactions. The dataset type effect is also significant, and the interaction between dataset type and prompt category is substantial. These results indicate that prompt robustness is dependent on question type, prompt change, and model.
Blogger's Review: This study reveals the differences in robustness of large language models when addressing various question types, emphasizing the need to consider the nature of questions when designing evaluations. This finding is crucial for model assessment and optimization, especially in political and social contexts. Understanding model performance across different scenarios will help us better utilize and improve these technologies.