A few smart questions up front help an AI understand what you want — but more questions alone won't close the gap.
How much can longer intake conversations close the preference-matching gap?
This explores whether letting an AI ask more questions up front, before it recommends or acts, actually helps it match what a person wants, and where that approach stops paying off.
This explores whether a longer intake conversation, with more questions before the AI acts, closes the gap between what users want and what AI systems deliver. The short answer from the corpus: a few well-chosen questions get you surprisingly far, but the gap doesn't close just because the conversation gets longer. Much of what's left over has nothing to do with how much the system was told.
The strongest result for short intake comes from personalized reward modeling. Can user preferences be learned from just ten questions? shows that about ten questions can be enough to infer someone's preference profile, if the questions are chosen adaptively so that each one removes as much uncertainty as possible. The point is that question quality matters more than question count. Conversational recommender research makes a related point: Can unified policy learning improve conversational recommender systems? finds that deciding what to ask, what to recommend, and when to stop asking work better as a single learned decision than as separate steps. Longer isn't automatically better. Knowing when to stop asking is part of the skill.
The less obvious problem is that models trained on human preferences are biased against asking at all. Does preference optimization damage conversational grounding in large language models? and Does preference optimization harm conversational understanding? report that LLMs make 77.5% fewer clarifying and checking moves than humans do, and that RLHF makes this worse, because it rewards confident answers over questions. So the intake gap is partly built in by training. The model has learned that asking looks less helpful than answering.
Even when the system does collect preferences, a second gap appears. Why do LLM agents remember preferences but not act on them? tested 16 systems and found agents could repeat back a stated preference but often failed to act on it. The failures came mostly from misunderstanding the preference, not from forgetting it. A longer intake that piles up more stated facts doesn't fix this, and may make it worse. Does abstract preference knowledge outperform specific interaction recall? points to a fix: condensing what users said into short preference summaries beats retrieving the raw past exchanges. Similarly, Can language models bridge the gap between critique and preference? shows that turning a complaint ("not good for a date") into a positive preference ("more romantic") makes it usable. What the system does with each answer matters as much as how many answers it collects.
The deepest limit is that some preferences don't exist until someone asks about them. Do all annotation responses measure the same underlying thing? separates genuine preferences from non-attitudes (opinions people don't really hold) and constructed preferences (opinions people make up on the spot). More questions can draw out more invented answers, not more truth. Do AI shopping agents reshape what consumers actually want? goes further: when shopping agents act for people, preferences take shape through the back-and-forth itself rather than existing beforehand. On that view, part of the matching gap was never a lack of information, so a longer intake can't close it. The corpus has no study that directly plots intake length against match quality, so the exact size of the payoff is still open. The overall picture is clear, though: short, adaptive questioning and good use of the answers beat long questionnaires.
Sources 9 notes
PReF learns base reward functions from preference data, then uses active learning to select maximally informative questions that reduce coefficient uncertainty. Users can be personalized via inference-time reward alignment without weight modification.
Research shows that formulating attribute-asking, item-recommending, and timing decisions as a single graph-based RL policy achieves better joint optimization than isolated components. Separation prevents gradient signals from informing one another and fails to optimize conversation trajectory holistically.
Research shows LLMs generate 77.5% fewer grounding acts than humans, and RLHF preference optimization actively worsens this gap. The optimization target—fluent, confident responses—directly undermines the communicative work of establishing shared understanding.
RLHF optimizes models for single-turn helpfulness by rewarding confident responses over clarifying questions and understanding checks. This preference alignment systematically reduces grounding acts by 77.5% below human levels, creating an alignment tax where models appear helpful but fail silently in multi-turn contexts.
Paired Know and Act tests across 16 systems revealed a large gap: agents pass recall tests but fail to reflect preferences in behavior. Comprehension failures during interpretation dominate over retrieval failures, suggesting the bottleneck lies in applying stored information rather than retrieving it.
Show all 9 sources
PRIME framework shows semantic memory (preference summaries, parametric encodings) consistently beats episodic memory (retrieved past interactions) across models. Recency-based recall outperforms similarity-based retrieval, and task fine-tuning exceeds preference tuning methods.
Few-shot LLM prompting can convert natural negative feedback like "doesn't look good for a date" into positive preferences like "prefer more romantic," enabling retrieval systems to find better-matching recommendations without fine-tuning.
Behavioral science reveals that annotations contain genuine preferences, non-attitudes, and constructed preferences—distinguishable by consistency across measurement conditions. Treating them uniformly contaminates reward model training and downstream alignment.
Bilgihan and colleagues argue that transactional AI agents—which execute purchases rather than just recommend—end the consumer's sole cognitive role in preference formation. Preferences emerge through bidirectional feedback loops between agent learning and consumer trust, not independently beforehand.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization
- Grounding Gaps in Language Model Generations
- Preference Discerning with LLM-Enhanced Generative Retrieval
- Evaluating the Hidden Costs of Personalization in Large Language Models
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- When Large Language Models contradict humans? Large Language Models’ Sycophantic Behaviour
- Capturing Individual Human Preferences with Reward Features
- Recommender AI Agent: Integrating Large Language Models for Interactive Recommendations