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

2026-07-08 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, long

Summary

A study by Sadia Kamal et al. investigates whether prompt robustness differs between objective questions with fixed answers and subjective questions asking for opinions or values in large language model evaluations. Evaluating four instruction-tuned model families (Gemma, Llama, Mistral, Qwen) across three objective datasets (MMLU, ARC, CulturalBench) and three subjective datasets (Political Compass Test, ValueBench, World Values Survey), the research applies various prompt changes, including wording, framing, and format, to measure answer consistency. Findings indicate that subjective datasets exhibit lower mean consistency (0.787) compared to objective datasets (0.849), with instability rising from 0.151 to 0.213. Option-order perturbations cause the most significant consistency drop, particularly for subjective questions, where Type-I consistency is 0.485 versus Type-II at 0.328. The study concludes that prompt robustness is not a uniform model property but depends on the model, dataset, dataset type, and specific perturbation category.

Key takeaway

For AI Scientists evaluating LLM behavior or inferring model values, recognize that prompt robustness is highly task-dependent. Your single-prompt survey measurements are notably fragile, especially for subjective questions. You should implement systematic robustness checks across various prompt forms, including wording, formatting, and option order, to ensure stable and interpretable results. This prevents misattributing prompt sensitivity to stable model beliefs.

Key insights

Prompt robustness in LLMs is task-dependent, with subjective questions being significantly less stable than objective ones under prompt variations.

Principles

• LLM prompt robustness is not a global property.
• Subjective questions are less stable than objective ones.
• Answer-presentation changes cause largest instability.

Method

The study used a binomial generalized estimating equation (GEE) to measure answer consistency across prompt variants, testing main effects and interactions of model, dataset, dataset type, and prompt category.

In practice

• Evaluate LLM values using robustness checks.
• Separate semantic from surface perturbations.
• Report consistency across prompt variants.

Topics

• Prompt Robustness
• LLM Evaluation
• Subjective Questioning
• Objective Questioning
• Prompt Perturbations
• Model Sensitivity

Best for: Research Scientist, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, Prompt Engineer

Related on AIssential

• Stability vs. Manipulability: Evaluating Robustness Under Post-Decision Interaction in LLM Judges — LLM judge decisions are stable under neutral conditions but highly reversible under targeted post-decision conversational challenge.
• Spurious Prompts: Can Irrelevant Prompts Steer Large Language Models? — Large Language Models exhibit surprising sensitivity to semantically irrelevant "spurious prompts," which can systematically steer their behavior.
• How reliable are LLMs when it comes to playing dice? — Current LLMs struggle with probabilistic reasoning, especially on counterintuitive problems and when faced with subtle biases or misleading prompts.
• Same Evidence, Different Answer: Auditing Order Sensitivity in Multimodal Large Language Models — Multimodal LLMs exhibit significant order sensitivity, with current prompt-level mitigations proving insufficient for general robustness.
• A classic brain test exposed AI's biggest weakness — Current LLMs struggle with sustained attention and cognitive control on demanding, longer-sequence tasks.
• Month in 4 Papers (April 2026) — LLM accuracy can unexpectedly improve with "rude" prompts compared to "polite" ones.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.