Generative AI and eXtended Reality (XR) have introduced to computer applications not only new challenges, but also the need to reconsider from a new perspective classical problems to which hardware and software systems may be subject to and their implications. For example, errors or disturbances in the communication channel between the physical and virtual worlds can lead to systems malfunctions and unintended output, on the one hand, but can also have implications on the creativity of the XR system, on the other hand. Indeed, in systems where Large Language Models (LLMs) mediate or generate responses, these processes can trigger novel and unexpected outputs, thereby representing affordances for creative interactions. In this paper, we explore this scenario by starting analyzing the robustness of LLMs to perturbations—in the form of scrambled character blocks of prompts—thus simulating a typical XR chat scenario where an avatar is driven by an LLM. As a preliminary study, we analyze the response behavior of two small LLMs using a limited sample size of 30 questions.
@inproceedings{braccini2026llm,
title = {On the LLM Robustness in a Simulated Conversational XR Scenario: A Preliminary Semantic Analysis},
author = {Braccini, M. and Aguzzi, G. and Baldini, P. and Roli, A.},
booktitle = {Image Analysis and Processing - ICIAP 2025 Workshops},
booksubtitle = {23rd International Conference, Rome, Italy, September 15–19, 2025, Proceedings, Part II},
year = {2026},
volume = {16170},
pages = {637--648},
publisher = {Springer Nature Switzerland}
}