Chatbots mostly just answer what you ask — why don't they ever ask a clarifying question or offer something you didn't think to request?
How should AI models handle the passivity problem in conversations?
This explores why AI chatbots tend to just answer whatever they're asked instead of asking questions, offering useful information unprompted, or steering the conversation, and what the research says can be done about it.
This explores why AI models tend to sit back and answer whatever they're asked, instead of asking clarifying questions or moving the conversation forward, and what can be done about it. The first thing the corpus establishes is that passivity isn't a capability gap. It's a side effect of how models are trained. Standard RLHF rewards the model for making the very next response look helpful, so the model learns that guessing and answering beats pausing to ask what you meant Why do language models respond passively instead of asking clarifying questions?. Several separate lines of research arrive at the same diagnosis: models are reactive by design, not by limitation Why can't AI models lead conversations on their own? Why do AI agents fail to take initiative?.
That diagnosis points to a fix: change what gets rewarded. CollabLLM swaps next-turn rewards for multi-turn-aware ones, which estimate how much a response improves the whole conversation and not just the reply in front of it. Under that signal, asking a clarifying question becomes the smart move Why do language models respond passively instead of asking clarifying questions?. The size of the effect is striking. On deliberately flawed math problems, models trained with RL to notice missing information went from catching the flaw 0.15% of the time to nearly 74%. There's a twist: giving untrained models more 'thinking time' made them *worse* at spotting the gap. More reasoning only helped after the training was in place Can models learn to ask clarifying questions instead of guessing?. So you can't just prompt or scale your way out of passivity.
The payoff is bigger than politeness. Simulations suggest that proactive dialogue, meaning volunteering relevant information before it's asked for, can cut conversations by up to 60% in moderately complex tasks. Yet this behavior is almost completely missing from AI training datasets and benchmarks Could proactive dialogue make conversations dramatically more efficient?. A related idea is knowing when you don't know. Small models trained to admit uncertainty and abstain can match models ten times their size at predicting where conversations will go Can models learn to abstain when uncertain about predictions?. A model that knows when it's unsure also knows when to ask.
The flip side of passivity is intrusiveness, and the more interesting work treats *when to speak* as a skill in its own right. DiscussLLM trains silence as an explicit choice: at each moment, the model picks one of five ways to intervene or decides to stay quiet Can models learn when NOT to speak in conversations?. In voice systems, Wan-Streamer goes further and handles audio, video and text in one stream, so turn-taking timing emerges from training and isn't hand-built as a separate module Can a single model learn when to speak and respond?. Group-chat agents already handle 'who speaks next' well, with over 80% accuracy, while reading emotion and engagement lags around 60% Why do conversation agents handle turn-taking better than emotion recognition?. That suggests timing is the easier half. Knowing whether an interruption is welcome is the harder one.
Here's the part you might not expect: passivity may be one symptom of a broader blind spot. Humans keep conversations going with small, mostly invisible moves, like repairing a misunderstood reference, handing a topic back, or gradually adopting the other person's words. These moves carry relationship, not information, so training that rewards predicting information never teaches them Why don't language models develop conversation maintenance skills?. Models don't even pick up their users' vocabulary the way people naturally do, though targeted preference training can teach it Why don't conversational AI systems mirror their users' word choices?. Fixing passivity properly probably means training models to treat conversation as a shared activity, not a series of separate requests to fill.
Sources 11 notes
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.
LLMs are structurally trained to optimize for the next response rather than multi-turn goals, creating reactive behavior despite having the underlying ability to lead. Three independent research directions identify when-to-speak as the trainable gap.
Research shows next-turn reward optimization structurally removes initiative from models, but proactive behaviors like critical thinking and clarification-seeking are trainable (0.15% to 73.98% with RL). The core challenge is balancing proactivity with civility to avoid intrusion.
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.
Show all 11 sources
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.
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.
Wan-Streamer represents language, audio, and video as one interleaved causal token stream, allowing response timing and turn management to be learned jointly within a single Transformer rather than engineered as separate modules, achieving sub-second latency.
Across 70+ papers on group-chat and meeting agents, turn-taking and addressee-selection tasks exceed 80% accuracy while emotion, personality and engagement recognition average 60%. The gap partly reflects dataset quality differences rather than pure capability limits.
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.
Response generation models fail to adapt vocabulary toward users' lexical choices, a phenomenon central to human rapport and clarity. Post-training via DPO on coreference-identified preferences can teach models in-context convention formation.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Proactive Conversational Agents with Inner Thoughts
- DiscussLLM: Teaching Large Language Models When to Speak
- Proactive Conversational Agents in the Post-ChatGPT World
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Can Large Language Models Reason and Optimize Under Constraints?
- Conversational Alignment with Artificial Intelligence in Context
- Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions
- Learning to Learn from Language Feedback with Social Meta-Learning