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

Technical Summary: Prompt Robustness Is Task-Dependent

Problem Statement
Current evaluations of Large Language Models (LLMs), particularly survey-style assessments used to infer model values, beliefs, or political stances, rely on a fragile assumption: that a single prompt yields a stable and meaningful measure of model behavior. Prior research indicates that minor variations in prompt format, wording, and answer presentation can significantly alter model outputs. This instability is critical for survey-based evaluations, where responses are often interpreted as evidence of a model's intrinsic social attitudes or political alignment. If a model's answer shifts when the underlying task meaning remains constant, the evaluation conflates task ability with prompt sensitivity. This paper investigates whether this sensitivity behaves differently for objective questions (with fixed correct answers) versus subjective questions (asking for opinions or values).

Methodology
The study employs a unified robustness testing framework across two distinct categories of datasets and four instruction-tuned model families (Gemma, Llama, Mistral, and Qwen).

• Datasets:
  • Type-I (Objective): MMLU, ARC, and CulturalBench-Easy. These consist of multiple-choice questions with fixed answer keys.
  • Type-II (Subjective): Political Compass Test, ValueBench, and World Values Survey (WVS) items. These consist of opinion-based statements or agreement scales without a single "correct" answer.
• Perturbation Taxonomy: The authors curated a set of prompt variations designed to preserve the intended task while altering surface features. These are categorized into:
  • Semantic Framing: Paraphrase, lexical substitution, spelling noise, and logical equivalence.
  • Surface Form: Format variation (spacing, casing, separators).
  • Answer Presentation: Label substitution (changing A/B/C/D to symbols/numbers) and option shuffling (reordering choices).
• Experimental Protocol:
  • All models were run with deterministic decoding (temperature 0) to isolate prompt effects from sampling noise.
  • Outputs were normalized to canonical labels (A, B, C, D) to ensure consistency across different prompt formats.
  • Metric: Answer consistency ($C_i$) was calculated as the ratio of the most frequent answer to the total number of valid prompt variants for a given item.
  • Statistical Analysis: The authors utilized binomial Generalized Estimating Equation (GEE) models to account for the non-independence of items appearing under multiple related prompt variants. They tested main effects and interactions for model, dataset, dataset type, and prompt category.

Key Results
The statistical analysis reveals that prompt robustness is not a uniform property but is highly dependent on the interaction between the question type, the specific perturbation, and the model.

1. Subjective vs. Objective Instability: Subjective datasets (Type-II) exhibit significantly lower consistency than objective datasets (Type-I). The mean consistency for objective datasets is 0.849, compared to 0.787 for subjective datasets. This gap is most pronounced under option order perturbations, where subjective consistency drops to 0.328 versus 0.485 for objective tasks.
2. Perturbation Sensitivity: Not all prompt changes affect models equally.
   • Option Order: This is the strongest source of instability across all models and datasets, reducing mean consistency to 0.407.
   • Semantic Perturbations: Paraphrase, logical equivalence, and lexical substitution are comparatively stable, with consistency scores generally exceeding 0.90.
   • Format and Labels: Format variation and label substitution cause moderate instability, particularly in subjective tasks.
3. Interaction Effects: The GEE models confirm significant interactions between dataset type and prompt category ($\chi^2 = 175.435, p < 10^{-35}$). This indicates that subjective tasks are not merely less consistent on average; they are uniquely sensitive to how answers are presented (e.g., order and labels) compared to objective tasks.
4. Model Variance: Robustness varies by model. While Qwen generally demonstrates higher stability under option shuffling than other models, all models show significant sensitivity to specific perturbation types. No single model is robust across all dataset types and perturbation categories.

Contributions
The paper makes three primary contributions:

1. Curated Perturbation Set: It synthesizes a broad taxonomy of prompt perturbations from prior work on prompt sensitivity, multiple-choice formatting, and survey bias, adapting them for a unified evaluation across objective and subjective domains.
2. Unified Robustness Test: It provides a direct comparison of prompt robustness between objective and subjective datasets, isolating general prompt noise from instability specific to value-based inquiries.
3. Statistical Evidence of Task-Dependence: Using clustered statistical tests, the study demonstrates that prompt sensitivity is a joint function of the model, the dataset, the dataset type, and the perturbation category, rather than a static model attribute.

Significance and Claims
The authors claim that their findings necessitate a shift in how LLM evaluations, particularly those inferring values or beliefs, are conducted.

• Fragility of Single-Prompt Surveys: A single prompted response to a political or value question should not be treated as a stable measure of a model's beliefs unless it survives prompt variation.
• Interpretation of Instability: For subjective tasks, prompt variation is not merely noise; it reveals the fragility of the measured stance. A model may use wording, label choice, or option order as cues for how to respond, meaning the "belief" is constructed by the prompt structure rather than retrieved from a stable internal state.
• Evaluation Best Practices: The paper argues that robustness should be reported as a relationship among model, dataset, and perturbation category, rather than a single global score. Researchers should separate semantic perturbations from surface perturbations to identify specific failure modes (e.g., instruction following vs. position bias) and must report consistency across prompt variants rather than relying on a single accuracy or survey score.

The study concludes that for value and belief evaluation, robustness to prompt variation must become a standard component of the evaluation design, as the current reliance on single prompts yields unstable and potentially misleading inferences about model behavior.