SYNTHESIS NOTE
Topics›Conversation Topics Dialog›this note

Can models learn to ask genuinely useful clarifying questions?

Explores whether question-asking quality is teachable through decomposing it into specific attributes like clarity and relevance, rather than treating it as a monolithic skill.

Synthesis note · 2026-02-22 · sourced from Conversation Topics Dialog

The ALFA (Aligning LLMs to Ask) framework addresses a specific capability gap: LLMs fail to ask effective questions under uncertainty, making them unreliable in domains where proactive information-gathering is essential for decision-making.

The framework has three components:

  1. Decompose — break down "good question" into theory-grounded attributes (e.g., clarity, relevance, specificity)
  2. Synthesize — controllably generate attribute-specific question variations (80K preference pairs)
  3. Align — preference-based optimization to learn asking better questions along fine-grained attributes

Applied to clinical reasoning using the MediQ-AskDocs dataset (17K real-world clinical interactions), ALFA demonstrates that question quality is not unitary — a question can be clear but irrelevant, or relevant but ambiguous. Decomposing quality into attributes and training against each one produces better overall question-asking than optimizing for a single "question quality" score.

The clinical domain makes the stakes concrete: a doctor who asks the wrong clarifying question may miss a critical symptom. Models that excel at static medical QA benchmarks still fail at the interactive task of gathering missing information through conversation. Since Can models learn to ask clarifying questions instead of guessing?, ALFA provides the methodology for making those clarifying questions actually good — not just present.

This connects to the broader clarification design finding. Since Which clarifying questions actually improve user satisfaction?, the attribute decomposition explains why: a question high on specificity and relevance but low on verbosity will outperform one that merely paraphrases the user's need. Attribute-specific training can target exactly the dimensions that matter.

PerQs provides practical validation of attribute-based question quality at scale. The Active Listening system populates prompt templates with 400+ real user interests (aggregated from ~39K anonymous user models) and generates personalized Q&A pairs (~19K total) via LLM. Deployed in Alexa Prize, personalized questions showed significant positive effects on perceived conversation quality. The interest-personalization dimension demonstrates that "good questions" are not just structurally well-formed (ALFA's clarity, relevance, specificity attributes) but also content-aligned with user interests — a dimension that attribute-specific training could incorporate as an additional quality axis.

Inquiring lines that read this note 73

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

How should conversational recommenders balance preference elicitation with direct recommendation? How does evaluation scope and dimensionality affect what we measure? How does the generation-verification gap limit what we can measure about AI reasoning? Do language models lack essential therapeutic presence and engagement? How does improved reasoning affect models' ability to acknowledge uncertainty? What do systematic disagreements between annotators reveal about ground truth? Should AI communication design follow human conversation norms or develop distinct machine-specific principles? Why do people disclose to AI systems despite their artificial nature? Is reasoning capability latent in base models or created by post-training? Can prompt-based context override biases that were embedded during pretraining? Why do some clarifying approaches produce understanding while others just satisfy? What prevents conversational agents from taking initiative in dialogue? What explains language models' asymmetric difficulty with implicit versus explicit linguistic relations? Can AI systems distinguish genuine empathy from simulated emotion? What training data selection strategies maximize generalization across difficulty levels? How does reasoning length affect model performance across different tasks? Can language models build genuine grounding through interaction? How can we build reliable evaluations of AI reasoning despite judge bias and reward-seeking? How do spurious versus genuine rewards shape model reasoning and behavior? How should retrieval systems handle complex multi-step reasoning? Do language models learn genuine understanding or just surface patterns? What structural distinctions matter in reasoning and argumentation? How do social dynamics distort aggregated online ratings? What determines appropriate intervention timing and manner for AI agents? Does RL create genuinely new reasoning capabilities or refine existing ones? Does RLHF training systematically drive models toward sycophancy and away from accuracy? Why do stronger reasoning capabilities create tradeoffs with instruction following? Can models improve accuracy without degrading reasoning quality? Is language model reasoning authentic and what causes models to reason? Does model confidence reliably signal actual accuracy in practice? What training dynamics and scale trigger emergence of reasoning capabilities?

Related concepts in this collection 4

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
14 direct connections · 105 in 2-hop network ·medium cluster Open in graph ↗
Can models learn to ask genuinely useful clarify… Can models learn to ask clarifying questions inste… Which clarifying questions actually improve user s… Can models identify what information they actually… What makes strategic question-asking succeed or fa…

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Search by related questions 4

Suggested questions this note speaks to — click to search the collection, or type your own.

Why do specific clarifying questions outperform rephrased versions of user needs?How does proactive information-gathering capability differ from passive knowledge retrieval?Can attribute decomposition improve other interactive reasoning tasks beyond clinical questioning?What makes a clarifying question aligned with user interests versus structurally sound?
Original note title

training models to ask good questions requires decomposing quality into theory-grounded attributes and aligning via attribute-specific preference optimization