Home // SECURWARE 2025, The Nineteenth International Conference on Emerging Security Information, Systems and Technologies // View article
Authors:
Salvatore Vella
Salah Sharieh
Alex Ferworn
Keywords: large language models; bias; threat.
Abstract:
Small Large Language Models (LLMs) are now integrated into devices we use every day, but their reliability under prompt variations remains understudied. We see them on cell phones and many other devices. We present a study of prompt variation in small LLMs, focusing on the effect of prompt formatting changes on multiple-choice reasoning tasks, even when the prompt provides the correct answer. We evaluate LLaMA-3 (1B and 4B), Google Gemma (1B and 4B), Alibaba Qwen (1.5B and 3B), Microsoft Phi-3 (4B), IBM Granite (2B) and the smaller OpenAI models (gpt-4o-mini, gpt-4.1-mini, gpt-4.1- nano) on the CommonsenseQA and OpenBookQA benchmarks. Our findings reveal that reordering of answer choices causes statistically significant performance drops, even when the correct answer is explicitly present in the prompt. For very small models, the results are dramatic. Statistical tests, including paired t-tests and McNemar’s test, are used to confirm the significance of the results. These results suggest that smaller LLMs rely on heuristics rather than reasoning, as they fail to grasp the correct answer even when it is explicitly provided. This prompt-order sensitivity, where providing the correct answer, is a unique attack surface in LLM systems, allowing adversaries to manipulate prompt structure to create errors. This work suggests additional testing is needed before deploying LLM-based systems.
Pages: 89 to 95
Copyright: Copyright (c) IARIA, 2025
Publication date: October 26, 2025
Published in: conference
ISSN: 2162-2116
ISBN: 978-1-68558-306-4
Location: Barcelona, Spain
Dates: from October 26, 2025 to October 30, 2025