Why do AI chatbots almost always just answer instead of asking 'wait, what do you mean?' first?
What training signals cause LLMs to prioritize direct answers over clarification?
This looks at why language models usually just answer, even when the request is unclear and a good human helper would ask a question first, and which parts of training push them that way.
This looks at why language models usually just answer, even when the request is unclear and a good human helper would ask a question first. The corpus points mainly at one cause: how training scores each reply. Standard RLHF rates a response on how helpful it looks right away, as a single turn. A clarifying question looks less helpful in that moment than a confident answer, even when it would make the whole conversation go better. The CollabLLM work names this directly: next-turn rewards teach models to respond passively rather than find out what the user actually wants. Rewards that estimate a reply's value across the rest of the conversation bring back active questioning Why do language models respond passively instead of asking clarifying questions?.
The surprising part is that the models can tell something is ambiguous. When asked directly whether a question is ambiguous, they often say yes. Then they answer it straight anyway. Giving them retrieved documents makes the gap wider, because the extra material makes them more confident instead of more cautious. That points to what training rewards, not to a limit in what the model understands Do LLMs recognize ambiguity but ignore it anyway?. The cost shows up across a conversation. In 200,000+ simulated conversations where instructions arrive a piece at a time, models lock onto early guesses and rarely recover. Accuracy falls by about 39% on average, and agent-style fixes win back only part of that loss Why do language models fail in gradually revealed conversations? Why do AI assistants get worse at longer conversations?.
The models do ask some questions, just about the wrong things. In collaborative Minecraft building dialogues, people skip questions about which object is meant, because shorthand is efficient. They do ask about the task itself, often after they've started acting. LLMs do the opposite: they over-ask about wording and under-ask about the actual job Why do humans and LLMs ask clarification questions differently?. So the training signal doesn't just suppress clarification. It leaves models with a poor sense of when a question is worth asking. A related idea treats vague queries as missing context: with little to go on, models fall back on a blended average of their training data and give a generic reply rather than asking for specifics Why do large language models produce generic responses to vague queries?.
The fixes in the corpus mostly change what gets rewarded or what the model practices. Social meta-learning turns ordinary tasks into teacher-student dialogues. The model only succeeds by drawing information out of a partner who knows more, so asking becomes a way to solve problems rather than a tone of voice Can LLMs learn to ask for feedback during problem solving?. Another system, ARIA, adds a step where the model checks its own uncertainty and sends clashing rules to a human. Its authors argue that some choices depend on context the model can't see, so asking is the only correct move Can LLMs learn reliably at test time without human oversight?. Even very small prompt changes help. Adding "wait a minute" before the model's reply, or telling it to correct false assumptions in a question, raised safety scores by up to 70% on one benchmark across six models Can simple discourse markers make LLMs question harmful assumptions?. That suggests the questioning behavior is already in the models and the training signal mostly keeps it switched off.
There's also a reason clarifying works so well once it happens. Models answer more accurately after a real question-and-answer exchange than after being handed a neatly rewritten, unambiguous version of the same request. Dialogue is closer to the conversations they saw in pretraining Do clarifying dialogues help LLMs answer better than rewritten inputs?. So the model learned from human conversation, where asking is normal, and later training then teaches it to stop asking.
Sources 10 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.
Models identify ambiguity when explicitly judged but overwhelmingly answer ambiguous questions directly. Retrieved context widens this gap by increasing confidence in answers rather than prompting clarification. Training incentives reward direct answers over ambiguity acknowledgment, not representational limits.
Across 200,000+ conversations, all major LLMs show 39% average performance drop in multi-turn settings due to locking into incorrect early guesses. Agent mitigations recover only 15-20% of this loss.
LLMs perform at 90% accuracy with single-message instructions but drop to 65% across natural conversation. Models lock into early guesses when information arrives gradually and cannot course-correct, a behavior induced by RLHF training that rewards helpfulness over clarification.
Analysis of 100 Minecraft dialogues shows humans rarely ask about referential ambiguity (treating it as efficient shorthand) but frequently clarify task-based uncertainty after acting. LLMs show the inverse pattern, over-clarifying referential issues and under-clarifying task concerns.
Show all 10 sources
Unlike social-media context collapse, which flattens multiple audiences, LLM collapse occurs when users provide insufficient contextual scaffolding and models default to blended training-data priors. This distinction suggests remedies should focus on query verification and user-driven context specification rather than platform controls.
Research shows that reformulating static tasks as pedagogical dialogues—where a teacher has privileged information and the student must learn to extract it—trains models to actively engage conversation as a problem-solving tool, not just imitate dialogue patterns.
ARIA demonstrates that LLMs can adapt during inference through three integrated components: structured self-dialogue for uncertainty assessment, timestamped knowledge bases for conflict detection, and human-mediated resolution queries. Autonomous systems fail at reconciling contradictory rules because the correct choice depends on context outside the system.
Two minimal prompting interventions achieve 70% improvement on Cancer-Myth and 40% on SAGE-Eval across six LLMs. The discourse marker 'wait a minute' appears to work by triggering epistemic vigilance rather than adding new information.
LLMs answer more accurately when given clarifying questions and answers than when given minimally-edited disambiguated versions of the same ambiguous input. The gap exists because clarifying exchanges match the model's pretraining distribution better than artificially constrained rewrites.
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
- Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- Referential ambiguity and clarification requests: comparing human and LLM behaviour
- Accommodation and Epistemic Vigilance: A Pragmatic Account of Why LLMs Fail to Challenge Harmful Beliefs
- Clarify When Necessary: Resolving Ambiguity Through Interaction with LMs
- QuestBench: Can LLMs ask the right question to acquire information in reasoning tasks?
- HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants