AI usually guesses instead of asking what you meant. Can it be taught to stop and ask first?
Can models learn to ask clarifying questions instead of answering prematurely?
This explores whether language models can be trained to notice when a request is missing something important and to ask about it, instead of guessing and answering right away.
This explores whether models can learn to notice missing information and ask about it instead of guessing. The corpus says yes, and that models don't do it by default because of how they are trained, not because they can't. Self-play is one route: STaR-GATE has a model fine-tune itself on the questions that led to better responses. After two rounds it was preferred over the base model 72% of the time, with no human-written question examples (Can models learn to ask better clarifying questions through self-improvement?). Splitting 'a good question' into parts is another. ALFA scores clarity, relevance and specificity separately and trains on 80K preference pairs. That beats a single overall score, most clearly in clinical reasoning, where the right follow-up question changes the decision (Can models learn to ask genuinely useful clarifying questions?).
The passivity comes from the reward signal. Standard RLHF rewards whatever looks most helpful in the very next reply, and a clarifying question looks unhelpful by that measure, so models learn to answer confidently. CollabLLM instead rewards the estimated value of the whole conversation, and that lets models actively work out what the user wants (Why do language models respond passively instead of asking clarifying questions?). The same training gap shows up in reasoning models. Given a question with a missing premise, they produce long, redundant chains of thought, while non-reasoning models often just say it can't be answered. Training taught them to produce reasoning steps but never to stop and say something is missing (Why do reasoning models overthink ill-posed questions?).
The most surprising finding is that asking can emerge without anyone training for it. Social meta-learning trains models on complete problems, recast as dialogues where a teacher holds the key information and the student has to draw it out. Those models then handled underspecified tasks by asking for what they needed and holding back an answer, even though they had never seen that kind of task (Can models learn to ask clarifying questions without explicit training?). The underlying lesson is that conversation is a tool for getting information, and that lesson also carries over to asking for feedback mid-problem (Can LLMs learn to ask for feedback during problem solving?).
The skill is fragile, though. In one study, reinforcement learning raised a model's rate of spotting deliberately flawed math problems from 0.15% to 73.98%. Letting an untrained model think longer made it worse at this, and only after RL did extra thinking help (Can models learn to ask clarifying questions instead of guessing?). More compute doesn't make a model notice what's absent unless it was trained to look for it.
Knowing to ask is only half of it. The other half is choosing which question. UoT imagines the possible answers to each candidate question and picks the one that would shrink the diagnostic uncertainty the most, which favors a specific question over a generic 'can you tell me more?' (How can models select the most informative question to ask?). Asking also sits next to abstaining. Small models trained to hold back when unsure matched models ten times their size on conversation forecasting, which suggests both skills are present but undertrained in standard LLMs (Can models learn to abstain when uncertain about predictions?).
Sources 9 notes
STaR-GATE iteratively finetunes a model on questions that increase response quality, achieving 72% preference over the base model after two iterations. The research shows preference elicitation is trainable through self-play without human question supervision.
The ALFA framework breaks down question quality into theory-grounded attributes (clarity, relevance, specificity) and trains models on 80K attribute-specific preference pairs. Attribute-specific optimization outperforms single-score training, especially in clinical reasoning where asking the right clarifying question directly impacts decision quality.
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.
Reasoning models generate redundant, lengthy responses to questions with missing premises while non-reasoning models correctly identify them as unanswerable. Training optimizes for producing reasoning steps but never teaches models when to disengage.
Models trained via SML on complete problems generalize to underspecified tasks by asking for needed information and delaying answers. The training paradigm instills a meta-strategy of using conversation as an information source, addressing the premature-answering failure mode.
Show all 9 sources
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.
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.
UoT combines uncertainty-aware scenario simulation with information-gain scoring and reward propagation to identify questions whose possible answers maximally reduce diagnostic uncertainty—providing a principled mechanism for specific, high-value clarification rather than generic prompts.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Learning to Learn from Language Feedback with Social Meta-Learning
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- STaR-GATE: Teaching Language Models to Ask Clarifying Questions
- Explain-Query-Test: Self-Evaluating LLMs Via Explanation and Comprehension Discrepancy
- QuestBench: Can LLMs ask the right question to acquire information in reasoning tasks?
- Aligning LLMs to Ask Good Questions A Case Study in Clinical Reasoning
- Can Large Language Models Reason and Optimize Under Constraints?
- Proactive Conversational Agents in the Post-ChatGPT World