TANDE: AI avatar study reveals young adults prefer nonverbal emotional cues

A team of researchers from multiple institutions has developed TANDE, an embodied conversational agent (ECA) powered by large language models (LLMs) specifically designed for emotionally charged dialogues with young adults. Published on arXiv and accepted for the 2026 ACM International Conference on Multimodal Interaction (ICMI), the study addresses a critical gap: how backchanneling modalities—the subtle verbal and nonverbal signals listeners use to show attention—affect user experience in AI-driven emotional support conversations.

In a within-subjects experiment with 36 young adults, TANDE tested two backchanneling conditions: nonverbal only (e.g., facial expressions, head nods) and combined verbal-and-nonverbal (adding spoken acknowledgments). Results showed a clear preference for solely nonverbal cues, which improved perceived rapport, empathy, and engagement. The findings also revealed gender differences, suggesting nuanced design requirements. The authors argue that ECAs for emotional well-being should prioritize subtle nonverbal feedback over verbose verbal interjections, especially for younger users who may be more sensitive to overbearing AI responses.