Effects of Interaction Modality on LLM-Based Information-Seeking

Abstract

Conversational Large Language Models (LLMs), increasingly also in voice mode, are reshaping how people seek and engage with information. Understanding modality effects on interaction patterns and user perspectives is essential for interaction design that balances factors such as trust, usability, and user experience. This paper presents a within-subject experiment (n = 33) comparing how text, voice, and mixed-modality interaction influence LLM-based information search, with participants composing recommendations in health-related scenarios. We find that voice-based and mixed-modality interactions were characterized by a more conversational interaction style, with higher turn ratios, more frequent follow-up questions, and shorter responses than text mode. User experience evaluations indicated differences in hedonic quality, but not in pragmatic quality. We observed no differences in the quality of participants’ recommendations, perceived cognitive load, trust, and anthropomorphism. Qualitative findings highlight situation-dependent preferences. We recommend that LLM-based search systems support mixed interaction modalities that allow for adjustment to preferences and task demands.