TL;DR
The same mathematical problem gets solved in different industries under different names, and you don't see it because you're stuck in your own field's terminology. The method makes an LLM rewrite the problem description with the method name and all industry vocabulary removed, leaving only "what is computed and how." This takes off the problem's "costume" and exposes its mathematical skeleton.
Ordinary search systems (including specialized scientific embeddings) latch onto topic and vocabulary, not the essence of the computation. An example from the paper: calculating a personalized antibiotic dose in medicine and the Kalman filter in control theory are mathematically the same problem, but they share not a single word of text and no citations. So search engines, embeddings and even human experts never connect them: they search by topic, not by structure.
The method is a single pass through an LLM: the model rewrites the problem description in "domain-neutral" terms (what is being estimated, from what objects, what the mathematical structure of the procedure is), and is explicitly forbidden from naming methods ("don't write 'Kalman filter'"; write "the stationary distribution of a damped random walk on a graph"). The model also adds a few abstract tags (type of computational structure, type of data, problem form: estimation/optimization/search, type of uncertainty). After this "undressing," plain text search finds hidden solution twins more accurately than advanced neural methods.
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.
Method Blueprint
STEP 1: Take the task/problem description → the LLM rewrites it,
removing method names and industry vocabulary,
keeping only "what is computed and how" (one prompt)
STEP 2: The LLM adds tags:
type of computational structure, type of data, problem form
(estimation/optimization/search/simulation), type of uncertainty
(in the same prompt)
STEP 3: Use the "stripped" description to search for analogs in other fields
(you can do it in the same chat: ask the LLM to suggest fields
that solve a structurally identical problem)
STEP 4: Cross-check the tags/structure of the matched candidates —
this is a filter that weeds out false matches by topic
Example
Task: A marketer is figuring out how to find the best of 5 ad creatives faster without blowing the whole budget on a long A/B test: every impression costs money, and you don't know the payoff in advance.
Prompt:
Describe my problem in terms of what is computed/estimated and how,
without words from my industry (marketing, advertising, conversion)
and without the names of known methods (A/B test, statistics).
Describe only: which objects are compared, which quantity
is maximized, and what logic governs the choice between "try
something new" and "use what already works".
My problem: I have 5 versions of an ad creative.
Every impression costs money. I don't know in advance which creative
gives the best conversion. I want to find the best version faster
without spending the whole budget on testing.
After the description, suggest 3-4 fields (not marketing)
that solve a problem with the same structure under a different name.
Result: The model first gives an abstract description like "allocating a limited number of trials among several options with unknown payoffs, balancing exploring new options against exploiting the best known one." Then it suggests fields: clinical trials (adaptive design, which patients get which treatment), casinos/slot machines (the multi-armed bandit problem), investment portfolio management. Then you can ask directly: "how is this problem solved in clinical trials? give me a concrete algorithm" and get a ready method (e.g., Thompson sampling) that transfers to marketing.
Why it works
People and LLMs think by topic, not by structure. If a text is about medicine, you look for solutions in medical papers: the brain (and the algorithm) sorts by vocabulary and context, not by the math inside. That's what trips us up: the same computational problem hides under a dozen different names in different "tribes" of science and business that never read each other.
The LLM's strength is that it can rewrite text in someone else's terms if you explicitly forbid the familiar words. Tell it "don't name the method, describe its math," and the model stops grabbing the first familiar label and is forced to break the problem into components: what the object is, what we're optimizing, what the search logic is.
The method uses this as a filter: forcibly stripping jargon and names makes both the model and the human see the bare skeleton of the problem. After that, comparing skeletons is trivial; there's no distracting noise from industry words.
Control levers: - The tag set (structure/data/problem form) can be replaced with your own; for a business context it might be "resource type," "constraint type," "decision horizon." - You can ask the model at first not to name any field at all, only pure math. This strengthens the "skinning" effect if the model slides back to familiar answers too quickly. - The strictness of the ban ("don't name methods" vs. "don't name anything specific to your field"): the stricter it is, the stronger the abstraction, but also the higher the risk that the model produces an overly generic description that's useless.
Prompt template
Describe the problem in terms of what is computed/estimated and how.
Don't use words from the {industry} industry or the names of known
methods ({method_examples}). Describe only:
- which objects are involved,
- which quantity is estimated or maximized,
- what the logic/structure of the solution procedure is.
My problem: {problem description}
After the description: suggest 3-4 fields far from {industry}
that solve a problem with the same structure under a different name.
For each one, briefly say what the method is called there.
What to fill in:
- {industry} — your field (marketing, HR, logistics)
- {method_examples} — ways of solving this problem you already know, which need to be "hidden"
- {problem description} — the problem itself, as concretely as possible
🚀 Quick start — paste into the chat:
Here is the template for finding solutions by "skinning" a problem.
Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask what field your problem comes from and which solution methods you already know, because without that there's nothing to "hide": the model needs to know which jargon and names to exclude so it doesn't slide back to the familiar answer.
Limitations
⚠️ Works only if an analog exists. If the problem really is unique, the method won't invent a solution: the model will offer its best guess, but with no guarantee that it has actually been solved somewhere.
⚠️ Sensitive to wording. If even a bit of jargon remains in the description, the model latches onto the topic rather than the structure. The author explicitly tested this and confirmed it with a "costume-swap" experiment on the problems.
⚠️ In the original, it's part of a larger search system. The researcher built a full retrieval pipeline (text search + tag filter) to sort hundreds of scientific papers. You don't need that for a one-off use: the principle "skin it → find the twin" works in a single conversation with an LLM with no infrastructure at all.
Resources
Eryk Kulikowski, LIBIS, KU Leuven. "Same Problem, Different Field: Cross-Domain Solution Import via Domain-Stripped Computational Fingerprints," JCDL 2026. The author releases the benchmark (109 papers, 18 "method families"), the code and the distillation prompt openly.