Task-specific dynamics of overlapping interactions in human–LLM chat

Determine, for each task category in human–LLM text-based conversation (including goal-oriented tasks such as question answering and text summarization and open-ended tasks such as social dialogue and metaphor generation), how overlapping interactions unfold in detail, specifying when and how overlap behaviors (e.g., preemptive answering and backchanneling) should occur to ensure effectiveness and appropriateness across tasks.

Background

The paper introduces OverlapBot, a chatbot that supports text-based overlapping behaviors (preemptive answering, backchanneling, and interruption handling) to emulate natural human conversational overlap. Through studies, the authors observe benefits in perceived communicativeness and speed but also note variability in suitability across contexts.

In the design insights, the authors specifically highlight that different task types—goal-oriented versus open-ended—may require different overlap strategies, and they explicitly state that understanding these task-specific interaction dynamics remains an open question requiring future research.