What pacing moves make conversational agents sound like active listeners?

Three researchers coded ten video recordings of human-to-human active listening using open coding, treating "strategic use of silence segments" as the unit of analysis, until the categories reached saturation and no new types emerged. The resulting taxonomy names five context-aware pacing strategies: Reflective Silence, Facilitative Silence, Empathic Silence, Holding Space, and Immediate Response, further operationalized into eight specific moves (Recognize, Reconfirm, Re-engage, Reposition, Reconsider, Resonate, Holding, and Resolve). The authors frame the core move as reconceiving silence: "silence, in this context, is not an empty void but a powerful communicative tool to show attention and interest."

Each strategy is triggered by a different conversational context rather than applied uniformly. Reflective Silence appears when a speaker shares experience or feeling, or indirectly seeks advice ("I've been unsure if I'm..."), signaling thoughtful processing and reinforcing trust. Facilitative Silence is used in moments of ambiguity, disruption, or topic transition, acting as a subtle prompt for the speaker to clarify or re-engage. Empathic Silence applies when speakers voice deep feelings, beliefs, or values; by staying silent the listener signals acceptance without judgment. Holding Space is reserved for moments of intense vulnerability, creating a non-directive, emotionally safe space for the speaker to process without interruption. Immediate Response, by contrast, applies no deliberate silence at all and is reserved for explicit information-seeking needs. The authors built these into an LLM-based conversational agent that classifies incoming user input and selects among the five strategies dynamically, rather than using one pacing style for an entire conversation.

This taxonomy sharpens the binary framing in When should AI systems choose to stay silent?, which treats the problem as a single speak-or-stay-silent threshold (DiscussLLM's silent token, Inner Thoughts' intrinsic-motivation score). Here, staying silent is not one decision but at least four functionally distinct ones — reflective, facilitative, empathic, and holding — each tied to a different relational trigger (confirmation, ambiguity, emotional depth, vulnerability) rather than a single relevance-or-not judgment. The companion note Can pacing strategies make conversational agents feel more human? reports how this same taxonomy performed when implemented and tested against a static-pacing baseline.

The taxonomy rests on a small formative sample — ten video cases — coded qualitatively, which is appropriate for distilling categories but not for establishing their prevalence or completeness across populations; the authors themselves caution against generalizing "our specific results or five-strategy framework to all populations," noting that user background and traits are critical factors. The downstream system's strategy-selection accuracy of 86.2% also implies that roughly one in seven live classifications picks the wrong strategy for the context, a gap the excerpt does not resolve. The framework should be read as a validated descriptive structure for what pacing-as-communication can look like, not yet as a robust, population-general classifier.