When an AI isn't sure what you really want, should it ask — instead of just guessing and answering confidently anyway?
How can systems use dialog to resolve uncertainty about user intent?
This looks at how AI systems can talk with users to work out what they actually want, instead of guessing silently or answering with false confidence.
This looks at how AI systems can talk with users to work out what they actually want, instead of guessing silently or answering with false confidence. The corpus suggests the hard part isn't technical. Today's language models are rarely built to treat a conversation as a way of collecting information. They're trained to answer whatever is in front of them. Why can't conversational AI agents take the initiative? argues this passivity is built in: models learn to respond, not to lead, and their fluent answers hide how often they're guessing. Why do language models respond passively instead of asking clarifying questions? traces part of the cause to how models are rewarded. When feedback scores only the next reply, a clarifying question looks less helpful than a confident answer, even if that answer is aimed at the wrong goal. Rewarding the value of the whole conversation makes asking worth it again.
Older dialogue research saw this problem long ago. Spoken assistants misheard 15 to 30 percent of words in noisy settings, so designers couldn't assume they knew what the user said, let alone what they meant. Why do dialogue systems need probabilistic reasoning? describes systems that held a ranked set of possible user goals and narrowed it down turn by turn. That's a useful model for today's chatbots, which often settle on one reading of an ambiguous request too early. Can dialogue systems track both speakers' beliefs across turns? takes the idea further: it tracks what *both* sides believe and models how a conversation moves from partial to shared understanding, as in a doctor-patient exchange. Can command generation replace intent classification in dialogue systems? offers a practical alternative to sorting each message into a fixed list of intents. Instead, the system writes small commands that take the conversation's context into account.
The more surprising material comes from conversation analysis, the study of how people actually talk. Humans have names for two different repair moves. A clarifying question asked *before* answering is an "insert-expansion," and When should AI agents ask users instead of just searching? uses that idea to decide when a tool-using agent should stop and check with the user instead of quietly chaining searches away from what they meant. The other move happens *after* a wrong answer has revealed the misunderstanding ("no, I meant…"). Can AI systems detect and correct misunderstandings after responding? finds this "third position repair" is almost absent from current AI. Systems struggle to see that their earlier assumption was wrong and update it.
Knowing *when* to ask matters as much as knowing how. Can models learn to abstain when uncertain about predictions? shows that small models trained to know when they're unsure, and to hold back, can match models ten times larger. That kind of self-awareness is what should trigger a clarifying question. Can dialogue planning balance fast responses with strategic depth? uses a similar signal: it replies quickly when confident and switches to slower planning when unsure. Dialogue can also cut the work, not add to it. Could proactive dialogue make conversations dramatically more efficient? finds that offering relevant information without being asked can cut conversations by up to 60 percent. How can proactive agents avoid feeling intrusive to users? adds a warning: an agent that probes at the wrong moment, or overrides what the user said, feels intrusive. Good timing and respect for the user's control count as much as accuracy.
The takeaway: resolving intent through dialogue is less about the model's language skills and more about incentives and timing. Models need to be rewarded for the payoff over a whole conversation. They need to know when they're uncertain. And they need the two repair moves people use without thinking: ask before acting, and fix things gracefully after getting it wrong.
Sources 11 notes
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.
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.
Real-world speech recognition achieves 15-30 percent error rates in noisy environments, making deterministic flowchart dialogue systems unworkable. POMDP-based systems handle this by maintaining belief distributions over user intent rather than committing to single interpretations.
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.
Rasa's dialogue understanding architecture generates domain-specific commands instead of classifying intents, eliminating annotation requirements, handling context naturally, and scaling without degradation—treating understanding as pragmatics rather than semantics.
Show all 11 sources
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.
Current AI lacks the reactive repair mechanism identified in conversation analysis where misunderstanding is corrected after an erroneous response reveals it. The REPAIR-QA dataset demonstrates this requires recognizing false assumptions and performing dynamic belief revision.
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.
A framework combining a neural policy model (System 1) for familiar contexts with MCTS planning (System 2) for novel scenarios, switching based on the model's own uncertainty estimates, matches or exceeds pure MCTS performance while reducing computational cost.
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.
Intelligence and adaptivity alone create socially blind agents that interrupt poorly and override user direction. The Intelligence-Adaptivity-Civility taxonomy shows civility—respecting boundaries, timing, and autonomy—is essential to making proactivity welcome rather than intrusive.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Proactive Conversational Agents in the Post-ChatGPT World
- DiscussLLM: Teaching Large Language Models When to Speak
- Proactive Conversational Agents with Inner Thoughts
- Rethinking Conversational Agents in the Era of LLMs: Proactivity, Non-collaborativity, and Beyond
- Probing the Multi-turn Planning Capabilities of LLMs via 20 Question Games
- Plug-and-Play Policy Planner for Large Language Model Powered Dialogue Agents
- POMDP-based Statistical Spoken Dialogue Systems: a Review