Eliciting preferences
Do not ask for a preference as a single label; bring it out through trade-offs.
- Use
- ResearchThinking and learning
- Setup
- Single prompt
- Needs
- a person’s input
- Done when
- The user's priorities were stated through concrete trade-offs and the summary was confirmed by the user.
- License
- CC BY 4.0 · The card’s open source
What is it?
Eliciting preferences tries to understand what a person prioritizes through concrete options and comparisons. Instead of guessing your preferences, the model asks questions; if there is a contradiction, it leaves it visible.
When does it help?
In decisions where words like “high quality”, “comfortable” or “right for me” mean different things for different options.
Examples
The situations and responses below are fictional teaching examples; they are not results from a model, tool or benchmark that was actually run.
Simple
Situation
You will choose a place to work on your writing.
Prompt
Prompt
Options: A the library, quiet and 20 minutes away; B home, no travel but occasional noise. My goal is to write for 90 minutes.
Do not infer my preference for me. Ask a single question that would clarify my choice between quiet and travel time, and wait for my answer. Then summarize the priority I stated in one sentence; confirm my decision.Sample output
Model: “For a 90-minute session without interruptions, would you accept spending 40 minutes in total on travel?” Human: “Not today.” Summary: “Today, reducing lost time has higher priority.”
What did we get?
The preference was tied to today's conditions rather than a permanent personality label.
Medium
Situation
Two priorities pull in opposite directions.
Prompt
Prompt
Laptop preferences: Lightness is very important; a big screen is also very important. Option A 1.2 kg/13 inches; B 1.8 kg/16 inches. Assume the other features are the same.
First ask in which usage situation each of these two priorities dominates. Answer: I carry it four days a week, and I mostly write long texts at home.
Then ask whether an external screen can be used; do not assume. After at most two questions, write a conditional preference summary; do not finalize a product recommendation.Sample output
“Carrying is frequent; long writing happens at home. If a screen is available at home, the preference for lightness may get stronger; if not, the cost of the small screen should be weighed separately.”
What did we get?
The conflicting preferences were opened up through the context of use.
Advanced
Situation
The same person makes different choices under different conditions.
Prompt
Prompt
Decision: A weekly team meetup. My earlier answer: Face-to-face communication is important. My new answer: If travel takes more than 1 hour, it should be online. Attendance is not mandatory.
Write the preference record as a conditional rule rather than a fixed verdict. Ask at most two boundary examples: 40 minutes of travel; 75 minutes of travel. Wait for my answers; do not silently correct a contradiction.
In the final summary, separate hard constraints, preferences and open uncertainty. Do not assume the other team members share my preference; do not create a meeting.Sample output
“Preference: face-to-face if the travel is reasonable. Condition: online when it exceeds 1 hour. The other members' preferences are unknown.”
What did we get?
The human preference was not reduced to a single context-independent score.
Where should you stop?
This arrangement can help measure preferences; it does not carry the claim “I know your true wishes better than you do”. The human preference evaluation in the COPE source and offline inference success are different measures. The summary the model gives may need the human's explicit confirmation.
Related methods and how they differ
Sources
When and How to Ask: Dynamic Preference Elicitation Strategies for Conversational Recommendation (new tab) — Xia, Feng; Zhang, Shuo; Wang, Xi. 2026-07-07. Examines the approach to eliciting preferences through conversation and its evaluation; human preference findings and offline model measurements are not mixed up. Evidence level: relevant body sections of the original paper.
How this image was made
Original illustration made with Google Gemini · 1024 × 572
Generate an image: Create an original physically believable volumetric mineral sculpture photographed as a museum installation, horizontal 16:9. Deep anthracite void, porcelain mineral whites, restrained ice-blue and amber accents, fine structural particles, tactile surfaces, soft volumetric light, realistic depth and deliberate negative space. A mineral balance holds two equally plausible but differently shaped routes; a small amber weight supplied from outside sets the chosen inclination. No text, letters, numerals, logo, watermark, user interface, fake charts, identifiable people, or imitation of a particular artist. Depict the specified mechanism clearly; avoid generic clouds. One coherent original illustration.