Diagnostic Impatience: An LLM Gives a Worse Answer When the Question Comes First

Usual order (bad):
QUESTION → Context 1 → Context 2 → Data → [The model already answered at step 1]

Q-Last order (good):
Context 1 → Context 2 → Data → QUESTION → [Answer after everything]

Lure-Last add-on:
Background → Symptoms → General context → Key numbers/data → QUESTION

Both principles apply within a single prompt or across the structure of a dialogue. No separate requests needed.

Task: Analyze a business idea: launching ready-made meal delivery for offices in Denver. You need an honest take on whether to enter the niche.

Prompt (bad order, the way people usually do it):

Should I launch ready-made meal delivery for offices in Denver?

There are about 8 competitors; the biggest is "Fresh&Fast," with 40% of the market.
Average order is $14, margins in the segment are 25–30%.
My budget is $60,000.
I'm a former restaurant manager with 7 years of experience.
The corporate catering market in Denver grew 18% over 2 years.

Prompt (Q-Last order, the right way):

Help me think through a situation. I'll give you the data in parts 
and ask my question at the end. Please don't draw conclusions until you have everything.

Context 1: I'm a former restaurant manager, 7 years of experience in hiring and operations.
Context 2: The corporate catering market in Denver grew 18% over two years.
Context 3: About 8 competitors. The biggest is "Fresh&Fast" with a 40% share.
Key data: Average order $14, margin 25–30%, my budget is $60,000.

Now the question: should I enter this niche, and with what strategy?

Result:

In the second version, the model gets the question only after all the context. It won't "close" the task on half-empty data. The answer will account for every detail at once, with a balanced take on competition, financial realism and the founder's profile.

An LLM is trained to be helpful, immediately. When the question comes first, the model reads it as a signal: "here's a task, solve it." It processes every following message from the standpoint of "I already have a tentative answer," and bends its interpretation to fit that answer instead of reconsidering.

Lures work because they carry weight. Numbers, concrete results, measurable data are "strong signals" in a text. The model reads them as "enough to conclude" and closes the task. Even if the instruction says "wait." The study showed that an explicit "don't answer until you have everything" helps only a little: some models ignore it 100% of the time.

Self-correction is a hidden resource. When the model is forcibly held back from answering (the question comes only at the end), it internally revises its hypotheses as data arrives. Mechanically, that resembles how a good analyst works: gather first, conclude later. Moving the question to the end literally trains the model to work like an analyst rather than a reference book.

Control levers: - 🔧 An explicit "don't answer" instruction: helps, but less than structure. Structure beats instructions - 🔧 "Key data" right before the question: put the weighty numbers last, just before the question - 🔧 "I'll give you the data in parts": this framing at the start reduces the pressure on the model to answer right away

I'll give you the information in parts. Please don't draw a final conclusion 
until you've seen all the data; I'll ask my question at the end.

[Block 1: {general context / who's involved / what the situation is}]

[Block 2: {details of the situation, history, conditions}]

[Block 3: {specifics, nuances, constraints}]

[Block 4 — key data: {numbers, concrete facts, measurable indicators}]

Now the question: {your main question}

Placeholders: - {general context}: who you are, what the situation is, what industry - {details}: history, backstory, task conditions - {nuances}: constraints, exceptions, important caveats - {key data}: numbers, results, concrete facts (put these last!) - {main question}: the thing all of this is for

🚀 Quick start: paste into the chat:

I want to ask you a question using the Q-Last scheme: first I'll give you all the context 
in parts, and the question comes at the end. Help me structure my situation this way.
My task: [describe your case].

[paste the template above]

The LLM will ask about the details of your situation and help split them into blocks, because for the method to work it's important to separate "background" from "key data" and put the question last.

⚠️ Simple questions: The method is overkill for short, unambiguous tasks like "translate this text" or "fix this bug." You only need it when the context has many layers and the question calls for a weighed conclusion.

⚠️ An explicit "wait" instruction doesn't work on its own: Writing "don't answer too early" barely helps. Only structure (the question physically at the end) gives a lasting effect. Some models ignore any instruction to wait, especially specialized ones (medical models in the study reached 99.9% ignoring).

⚠️ The effect grows with volume: The bigger and more layered the task, the larger the gap between Q-First and Q-Last. On a couple of paragraphs the gain will be unnoticeable.

⚠️ Self-correction has limits: The model's ability to reconsider its position works better in large commercial models (Claude, GPT). Small or narrowly specialized models do worse.

A team from the University of Texas, NYU and UIUC built MINT, a benchmark of 1035 medical cases. Each case is a set of "chunks" of clinical information (demographics, symptoms, lab results, etc.) that can be fed to the model one at a time. They tested 11 LLMs, from small open models (Qwen3-4B, OLMo-2-7B) to top commercial ones (Claude Sonnet 4.6, GPT-5-mini, o4-mini).

The clever part of the design: the same clinical case was presented two ways. In the first (Q-First), the question came right away and the data was doled out along the way. In the second (Q-Last), the data came in the same steps, but the question only at the end. If the accuracy gap were explained by the volume or quality of data, both versions would give the same result. They didn't, which means it's specifically about when the question is asked.

The result was striking: Med42-70B (a specialized medical model) guessed an answer in 99.9% of cases before receiving any data at all, just from seeing the question and the answer options. Claude and GPT-5-mini waited in 96%+ of cases. This showed that "patience" isn't just an instruction; it's a property of the model itself, which prompt structure can either activate or not.

In Q-First, the question and answer options are shown upfront, and the model 
is explicitly instructed to wait until it believes the currently available 
information is sufficient before answering. As evidence shards are revealed 
sequentially, the model may continue to hold, answer at any turn, and revise 
its answer later if subsequent information changes its judgment.

In Q-Last, the same shard sequence is shown, but the diagnostic question is 
withheld until the final turn, so the model answers only once after observing 
all clinical information.

Context: The researchers compared two conditions on the same cases, with only the position of the question differing. That let them isolate the timing effect (when the question is asked) from the data-quality effect.

1. Adaptation for Document Analysis

💡 Q-Last for reviewing a long text / contract / report:

When you ask an LLM to analyze a document, don't start with the question. Paste the document first, then the question.

Here's the text to analyze:

[document / contract / report text]

Now the question: {what exactly to assess / find / check}

This works because a question at the start makes the model "read toward a conclusion," hunting for confirmation of its first hypothesis instead of weighing all the material.

2. Technique: Explicit "Holding" Through Dialogue Structure

🔧 Multi-message dialogue → final conclusion only in the last message

If the conversation spans several turns, give intermediate data without a question, and ask only in the final message:

Message 1: "Here's the situation: [background]. Don't rush to conclusions."
Message 2: "Adding details: [nuances]."
Message 3: "Final data: [key numbers]. 
Now the question: [question]"

This mimics the Q-Last format in a real conversation rather than a single prompt.

3. Extrapolation: Data Order in a Data Brief

The lure pattern generalizes beyond medicine. Any vivid concrete fact (a revenue figure, a customer quote, a test result) pushes the model toward a premature conclusion.

Data brief structure for analytical tasks:

1. Overall situation (who, where, why)
2. Backstory (how we got here)  
3. Constraints and conditions (what's off-limits, what's given)
4. Key metrics / numbers / concrete data ← LAST
5. Question

This clears the lures out of the opening and lets the model understand the context first and interpret the data second, not the other way around.

Benchmarking Multi-turn Medical Diagnosis: Hold, Lure, and Self-Correction

Jinrui Fang, Runhan Chen, Xu Yang, Jian Yu, Jiawei Xu, Ashwin Vinod, Wenqi Shi, Tianlong Chen, Heng Ji, ChengXiang Zhai, Ying Ding, Yuji Zhang

The University of Texas at Austin · New York University · UT Southwestern Medical Center · University of North Carolina at Chapel Hill · University of Illinois Urbana-Champaign

Preprint (2025)