Prompt Robustness Is Task-Dependent: Comparing Objective and Belief-Style Questions in LLM Evaluation

2026-07-06

AI Summary

This study investigates whether the stability of large language model (LLM) responses under minor prompt perturbations varies by question type—specifically, objective factual versus subjective belief questions. By applying multidimensional prompt perturbations (pertaining to wording, framing, and formatting) to four prominent instruction-tuned models across six benchmark datasets, and quantifying response consistency using binomial generalized estimating equations (GEE), the work provides the first empirical evidence that prompt robustness critically depends on the interaction between question type and perturbation method. The findings reveal significantly lower response consistency for subjective questions compared to objective ones, challenging the common assumption that LLM outputs directly reflect their internal beliefs or values, and highlighting the pronounced sensitivity of current models to prompt variations in value-laden tasks.

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 ask whether prompt robustness differs between objective questions with fixed answers and subjective questions that ask for 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 gives 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 large. These results show that prompt robustness depends on the question type, the prompt change, and the model.

Problem

• prompt robustness
• task-dependence
• objective questions
• subjective questions
• large language models

Innovation

• prompt robustness
• task-dependence
• objective vs. subjective questions
• large language models
• survey-style evaluation