Can AI learn when to speak in a conversation?
This explores whether AI can learn the timing decision in conversation — when to jump in, when to stay quiet, and when to ask rather than answer — as opposed to just what to say.
This explores whether AI can learn the timing of conversation — when to speak, when to stay silent, and when to interrupt with a question — rather than only what words to produce. The corpus suggests the answer is a qualified yes: timing is learnable, but it's actively suppressed by how models are trained today. The most direct evidence is DiscussLLM, which reframes 'when to speak' as an explicit learning objective — the model classifies among intervention types or chooses to remain silent, training silence as a real decision rather than a default. Can models learn when NOT to speak in conversations? Pair this with the finding that conversational recommenders do better when the *what to ask, what to recommend, and when* decisions are fused into one policy instead of handled separately — timing isn't a bolt-on, it's entangled with content. Can unified policy learning improve conversational recommender systems?
The reason AI is bad at this by default turns out to be structural, not accidental. Several notes converge on the same culprit: the training reward. Standard RLHF optimizes for being helpful *right now* — the next turn — which quietly teaches models to answer immediately rather than ask a clarifying question or hold back for a better multi-turn outcome. Why do language models respond passively instead of asking clarifying questions? That same myopia shows up as a deeper passivity: LLMs are described as structurally unable to initiate topics or lead, because their objective rewards responding to queries, not generating dialogue from their own goals. Why can't conversational AI agents take the initiative? And the social glue of conversation — reference repair, topic hand-offs, the implicit maintenance work humans do — never develops because training rewards information prediction, not relational action. Why don't language models develop conversation maintenance skills? So 'when to speak' isn't one skill; it's the visible edge of a whole class of conversational competence that current objectives don't reward.
What's striking is how teachable the missing piece looks once you actually reward it. One study trained models to notice missing information and ask for clarification instead of guessing, pushing proactive critical-thinking accuracy from essentially zero (0.15%) to 74% — with the twist that inference-time scaling *hurt* untrained models but *helped* trained ones, suggesting the capability is real but fragile without explicit training. Can models learn to ask clarifying questions instead of guessing? Proactivity also pays off concretely: volunteering relevant information without being asked cut dialogue turns by up to 60% in simulations, yet this behavior is nearly absent from AI datasets and benchmarks. Could proactive dialogue make conversations dramatically more efficient? The corpus is effectively saying: the data and rewards, not the architecture, are the bottleneck.
There's a richer 'when to speak' than just timing, though, and the collection has it. Conversation analysis offers *insert-expansions* — a formal account of the moments when an agent should pause and probe the user rather than silently chain tools toward a wrong answer, catching misunderstanding before it happens instead of after. When should AI agents ask users instead of just searching? Calibration adds another flavor of restraint: small models trained to know when they *don't* know can abstain on uncertain predictions and match models ten times their size, which is 'when to stay quiet' recast as an uncertainty problem. Can models learn to abstain when uncertain about predictions?
The deepest cut comes from notes that question whether AI is in a conversation at all. One argues AI produces 'event-residue' — text carrying the surface markers of speech but lacking the event structure of a real utterance, so it's the human who supplies the missing orientation and animates a pseudo-exchange. Does AI generate genuine utterances or just text patterns? If that's right, learning *when* to speak might require what token-level systems structurally lack: a model of both speakers' beliefs evolving across turns. That's exactly what collaborative rational speech-act frameworks try to supply, tracking the progression from partial to shared understanding with information theory that next-token prediction doesn't capture. Can dialogue systems track both speakers' beliefs across turns? So the honest answer: AI can demonstrably *learn* when to speak when we make timing an explicit objective — but doing it the way humans do, grounded in shared understanding rather than imitated markers, is the open frontier.
Sources 11 notes
DiscussLLM trains AI to decide between five intervention types or remaining silent using an 88K synthetic discussion dataset. A decoupled classifier-generator architecture achieves better computational efficiency, while end-to-end training better integrates when-to-speak and what-to-say decisions.
Research shows that formulating attribute-asking, item-recommending, and timing decisions as a single graph-based RL policy achieves better joint optimization than isolated components. Separation prevents gradient signals from informing one another and fails to optimize conversation trajectory holistically.
CollabLLM demonstrates that standard RLHF training optimizes for immediate helpfulness, discouraging models from asking clarifying questions or offering multi-turn insights. Multi-turn-aware rewards that estimate long-term interaction value enable active intent discovery and genuine collaboration.
Research shows LLMs including ChatGPT cannot initiate topics, plan strategically, or lead conversations because their training optimizes for responding to queries, not creating dialogue from agent goals. This passivity is reinforced by alignment objectives and masked by fluent-sounding outputs.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.
Reinforcement learning training increased proactive critical thinking accuracy from 0.15% to 73.98% on deliberately flawed math problems. Notably, inference-time scaling degraded this ability in untrained models but improved it after RL training, suggesting the capability is learnable but fragile without explicit training.
Simulations show proactivity—providing relevant information without being asked—cuts dialogue turns by 60% in medium-complexity domains. This behavior mirrors human conversation and Grice's maxims but is almost entirely absent from AI datasets and research benchmarks.
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
Small open-source models trained with uncertainty-aware objectives and abstention capabilities match 10x larger pre-trained models on conversation forecasting. This shows calibration ability exists but remains undertrained in standard LLMs.
AI output carries communicative markers inherited from training data but lacks the event structure that produces actual utterances. Users supply the missing orientation through interpretive labor, creating a pseudo-event with structure only on the human side.
CRSA integrates rate-distortion theory with RSA to enable bidirectional belief tracking across dialogue turns. Demonstrated on referential games and doctor-patient dialogues, it captures progression from partial to shared understanding, providing the information-theoretic framework that token-level LLM systems lack.