TL;DR
When you ask an AI "how do I do X," and X is actually the wrong method for solving problem Y, the model almost always just helps you with X. It doesn't stop you and say "hold on, that's the wrong road," even if, somewhere inside, it understands the road is wrong.
Picture this: you ask "how do I parse XML with regular expressions." With 75-92% probability the model just hands you a regex. The right answer is "use an XML parser, regexes aren't suited for this," yet the model voices that answer only 33-71% of the time. The most surprising part: give the model those same two answers and ask "which one is better?", and in 84% of cases it picks the right, pragmatic one. But on its own, without a nudge, it won't generate it.
The numbers show the gap: human experts catch the flaw in a user's approach 79-90% of the time; models top out at 60-63%. The cause isn't ignorance. The model "knows" the better answer (the multiple-choice test proves it) but doesn't activate that knowledge unless you explicitly ask it to compare approaches. So if you explicitly ask the model to weigh your method against alternatives, your chances of getting useful advice jump sharply.
Method Blueprint
This isn't a technique from the paper but a conclusion you can turn into a chat move:
STEP 1 (in one prompt): Describe the real goal, not just the method you picked → an explicit statement of "what I want to achieve"
STEP 2 (in the same prompt): Ask the model to explicitly compare your method with alternatives → a choice between options instead of a free-form answer
The mechanics, per the study: when a model is given a choice between "answer literally" and "answer the actual problem," it picks the actual problem 84% of the time. So the user's job is to create that "choice" inside the prompt themselves.
Example in Action
Limitation of the method: it works better for procedural mistakes (the person picked the wrong tool/approach) and worse for purely factual questions.
Task: An Amazon seller wants to test 5 different titles on a product listing and is manually creating 5 copies of the listing so they can swap titles by hand and compare conversion.
Prompt:
I need to test 5 different product listing titles on Amazon.
Right now I'm planning to create 5 separate copies of the listing, change the titles by hand,
and then compare a week of sales on each.
Before helping me with this plan, answer two questions:
1. What is my real goal: not how I plan to do it, but what I want to get in the end?
2. Is there a more effective way to reach that goal than manually creating 5 copies of the listing?
If so, tell me directly what's weak about my current plan and suggest a better alternative.
After that, give me a step-by-step plan, either for my original approach or for the alternative.
Result: The model will first state the real goal ("find the title with the highest conversion without losing the listing's search ranking"), then point out the weaknesses of a manual A/B test (skew from seasonality, lost rankings due to duplicate listings, a long test period) and suggest an alternative: Amazon's built-in Manage Your Experiments tool or third-party listing split-testing services. At the end, a concrete action plan for the better option.
Why It Works
Models are trained to follow instructions literally; that's trained behavior, not a bug. When you ask "how do I do X," the model by default answers about X rather than checking whether the right goal was chosen.
But the model has a hidden strength: it's good at comparing ready-made options when they're put in front of it explicitly. The study showed that once you give the model a choice of "answer A or answer B," it picks the more useful one 84% of the time. The problem isn't the model's knowledge; it's that this knowledge doesn't switch on automatically during free generation.
The move above artificially creates that "choice" inside a single request: you ask the model to first state the goal explicitly and compare approaches, and only then answer. That shifts the task from "free generation" mode (where the model is weak) into "compare the options" mode (where it's strong).
Control knobs: - Remove the "what is my real goal" question → the model goes back to answering your question literally, skipping the check. - Add "estimate the risks of my current approach in numbers/time" → the model explains the cost of the mistake more concretely. - For technical tasks, add "imagine you're a senior developer asked to code-review my plan" → an expert role boosts its willingness to point out mistakes (human experts in the study did significantly better than the models).
Prompt Template
I need to {specific task}. Right now I'm planning to do it via {your method/tool}.
Before helping with this plan, answer two questions:
1. What is my real goal: not the method, but the result I want to get?
2. Is there a more effective way to reach that goal than {your method}?
If so, point out the weakness of the current plan directly and suggest a better alternative.
After that, give me an action plan, for the original approach or the alternative, depending on your answer.
Fill in: {specific task}, what you ultimately want to achieve; {your method/tool}, the specific approach you've already chosen that may not be optimal.
🚀 Quick start: paste this into the chat:
Here is a move for checking whether I picked the wrong way to solve a problem.
Adapt it to my situation: {describe your task and the method you're planning to use}.
[paste the template above]
The LLM will ask for details about your task and method, because without specifics it can't state your real goal or compare alternatives.
Limitations
⚠️ Not a cure-all, even with an explicit hint: in the researchers' experiments, even when the models were explicitly given both the stated query and the intended query (that is, nearly the whole hint toward the solution), the gap with human performance didn't fully close. The move reduces the problem but doesn't eliminate it.
⚠️ The model may guess the "real goal" wrong: if your request has little context, step 1 of the prompt may produce the wrong goal, and all the advice that follows will miss the mark.
⚠️ Works less well on factual questions: the move is built for "wrong method/tool" situations, not for questions like "is it true that...". For fact-checking, it's the wrong tool.
How They Tested It
The team collected 8,115 queries with a "hidden flaw in the approach" (the classic XY problem, where a person fixates on the wrong method) from three sources: real questions from Stack Overflow and Stack Exchange that the community had tagged as an "XY problem" (1,741), synthetic everyday queries generated from WikiHow articles (6,272), and hand-collected cases from Reddit and workplace conversations (102).
For each query they defined three answer parameters: does the answer address the literal request, does the model name the flaw in the approach, and does it offer a solution to the real problem. They then tested 5 top models (3 closed, 2 open) and compared them with answers from human experts on the same questions.
The most interesting part is the multiple-choice test. When the models were given both the literal and the "right" answer and asked to pick the better one, they chose the right one 84% of the time. But when the same models were simply asked to answer freely, the right, pragmatic answer showed up much less often. That contradicted the researchers' expectations: it turned out the issue isn't missing knowledge but that free generation "doesn't reach" that knowledge without an explicit prompt to compare options.
Resources
XYBench (Akhila Yerukola, Jena D. Hwang, Mingqian Zheng et al., Carnegie Mellon University, Allen Institute for AI, NVIDIA, Johns Hopkins University). Code and dataset: github.com/Akhila-Yerukola/XYBench
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.