The paper asks how well large language models (LLMs) and people can combine familiar ideas in unfamiliar ways. It introduces KOMBINE, an open-ended benchmark that hands random pairs of unrelated, recognizable entities (283 entries across 21 domains, drawn from Wikidata) to solvers and asks for three kinds of combination: an association (a factual chain of triples linking the pair), an analogy (two paths with an identical relation sequence, used to invent a concept by projecting structure across the mapping) and a blend (a shared abstract 'generic space' plus a fused concept with structure from

Einstein called it “combinatory play”

Joining familiar ideas in unfamiliar ways is one of the engines of discovery. A new benchmark, KOMBINE , deals random pairs of ideas to 35 AI language models and 75 people — and measures how well each can link them, map one onto the other, and fuse them into something new.

1 Link — find a chain of true facts from one idea to the other.

2 Map — find a shared pattern, then carry an idea across it.

3 Fuse — find what both share, and blend them into one new thing.

Deal a pair — or tap any two dots — to see how far apart two ideas sit.

Each dot is one of the 280 ideas in the benchmark’s pool, from Gravity to Sushi. Nearby dots mean similar things to a language-embedding model (all-MiniLM-L6-v2, the one the paper uses to measure distance); the map was computed for this page. Numbers are true distances — the flat map only approximates them. The three combinations shown are example answers reported in the paper.

00 · Why measure this?

Big discoveries often join ideas that sat far apart.

Studies of science find that high-impact work tends to combine more distant ideas. Researchers increasingly hope AI models can help invent and discover. So: can they do combinatory play — and how do they compare with us? To answer, you first need a way to measure it.

“Combinatory play seems to be the essential feature in productive thought.” — Albert Einstein, quoted by Hadamard (1954), as cited in the paper

Einstein → violin → Mozart

Connect two distant things through a chain of real links. (Einstein played the violin; Mozart composed for it.)

immune system : body :: national defense : nation

Spot the same relational structure in two different worlds — then use it to invent.

organism + machine = cyborg

Fuse two concepts into a third with properties of its own.

These three forms are the main kinds of combinatorial creativity studied in cognitive science. KOMBINE tests all three on random pairs of ideas with no known connection, so answers can’t simply be remembered. Examples from the paper’s Table 1.

01 · Score

How do you score a creative connection?

Build a bridge from Radio to The Milky Way, one true fact at a time. Then compare routes. Every answer in the benchmark is judged on three things — and they don’t always agree.

Your route · arc height = size of the leap

Next link from Radio

Pick a first link. There are five ways out of Radio — one of them is false.

Distances computed for this page with the paper’s embedding model. The route network and the sample crowd were written for this page; the van Gogh route is the paper’s own example answer (Table 2a).

02 · Map

Carry an idea across.

An analogy lines up two worlds through the same relations. Here is one a model found between a blue whale and a mattress. First make the structure match — then use it to invent.

Step A · Lock the mapping — relation words must be identical

Mapping: 0 of 3 relation words match — not valid yet.

The benchmark is strict: both chains must use the exact same relation word at every step — no synonyms. That makes it shared structure, not loose resemblance. Numbers on the rungs: distance between the matched ideas (computed for this page); their average is the analogy’s surprise.

Step B · Invent by projection

Pick something from the mattress world that has no whale counterpart, and push it through the mapping.

Lock the mapping first.

Whale groomer drone — illustration from the paper (Table 2; images there are for illustration only), CC BY 4.0. The invention need not exist: inventing it is the point, as the solar system once did for the atom.

03 · Fuse

Fuse two into one — and find the bottleneck.

A blend isn’t two lists stapled together. It needs a shared abstract shape (the generic space), and at least one property that both inputs feed at once. The best blends also grow something true of neither.

3a · Sort a blend: Adam Smith + Bacteria = market colony

Where does each property of the market colony come from?

Market colony — illustration from the paper (Table 2; for illustration only), CC BY 4.0. Blend by claude-opus-4.5. The paper’s prompt gives another example: a liquid franchise (democracy + banking), whose one “allocates” slot carries both democracy’s votes and banking’s credit.

3b · Find the shared shape: Photosynthesis + Bread

Before a blend, you need one abstract description that is true of both . Three frontier models offered these. Which one fits both? Predict, then test.

Choose a description to lay it over both inputs.

Model answers and judgments from the paper’s Figure 4 and its caption; plain-language explanations paraphrased.

3c · The abstraction bottleneck

Most blends die at the first gate.

Models find a valid shared abstraction only 43.4% of the time. But once they have one, they describe a coherent blended concept 97.1% of the time. The authors call this the abstraction bottleneck . Each dot below is one blend attempt.

Average rates from §4.1. For a single model (or the human volunteers), Gate 1 = its blend validity and Gate 2 = coherent blends ÷ valid ones — derived from Table 10.

3d · Every pair was blendable by someone

19 frontier models × 19 pairs. Rose = invalid shared abstraction, white = valid. Tap a row or a model.

No row is entirely rose: for every pair, at least one model found a valid shared abstraction — no item was impossible.

Redrawn from Figure 4 of the paper (CC BY 4.0).

Why might blends be harder than analogies?

The authors’ explanation: an analogy describes one input using structure the other already has. A generic space must be a new description that compresses both inputs at once — likely more demanding.

They also suggest a lack of lived, sensory experience may play a role: in human brain-imaging studies, mental imagery and imagined experiences are linked to creative association and combining concepts.

04 · Scoreboard

Machines vs. people.

75 online volunteers took the same tasks — five pairs of each kind — scored by the same judges as the 35 models.

Predict first

Which task did people find easiest?

Share of answers that were valid on each task (§4.3).

Leaderboard · composite score, % of maximum

Scores combine validity, surprise and originality (plus, for analogy and blending, the invented concept’s quality), as a % of the maximum; failing validity zeroes an answer. The best model differs by task. Table 7 of the paper.

Humans scored below every model overall and below nearly every model on each task — mainly because they found valid connections less often. The volunteers were general online adults; the authors note that scientists or artists might score higher.

05 · Multiples

Great minds — or great models — invent alike?

In science, “multiples” are discoveries made independently by different people at about the same time. The authors looked for the AI version: different models, same pair of ideas, near-identical inventions.

Properties are (relation, object) pairs; similarity is the cosine similarity of their embeddings. Two properties count as “the same” when similarity ≥ θ; two inventions are a τ-multiple when at least τ properties match. Examples and similarities from the paper’s Table 4.

One in five inventions is a multiple.

Of 2,070 concepts invented by the 35 models, about one in five (19.9%) shares at least two near-identical properties with another model’s invention for the same pair. Blends converge 14× more than analogies; models from the same company 2.3× more than models from different companies. Each dot = 1%.

Same-company multiples, points above cross-company baseline (τ = 2)

Figure 6 of the paper (rates; 95% bootstrap intervals omitted here).

06 · Two minds

Different minds, different maps.

People scored lower — but they didn’t just give worse versions of the models’ answers. They reached into different parts of idea-space.

Each dot is one answer’s added words, placed in 2-D (principal components of the embedding). Red = humans, blue = models; filled = valid. On analogy and blending especially, the two groups occupy different regions. Figure 8 of the paper, CC BY 4.0, cropped per panel.

Which kinds of words each group used more

Percentage-point differences in how often each group used a word category (§4.3).

Models tend to name people and places and link them by authorship, study and transformation. People reach for everyday objects, states and feelings, linked by movement and plain being.

Word size = how much more one group used it. Figure 7 of the paper, CC BY 4.0.

07 · Back to the sky

Combinatory play, together?

KOMBINE gives a way to track progress toward AI systems — and human–AI teams — that can make the kind of distant combinations discovery runs on.

Today’s models already out-score typical volunteers, yet they stall at abstraction and often invent alike, while people reach different regions of idea-space. The authors suggest that pairing the two could widen the range of ideas. Combinatory play may work best as a team sport.

Deal yourself a pair from the benchmark’s pool — two ideas from different domains, just as the models and volunteers received them.

Evidence · How it was measured

The fine print, in plain words

Method

35 language models, each sampled at temperature 0.9.

Idea pairs drawn from 283 recognizable Wikidata entities across 21 domains (each with ≥ 78 Wikimedia sitelinks), always from different domains.

Factuality judged by Claude Haiku 4.5; analogy and blend quality by a panel of three AI judges (fair-to-good reliability, ICC 0.48–0.67).

A blind human rater agreed with the panel 75% of the time when the panel was unanimous; the ranking barely changes on those items.

Surprise and originality use the all-MiniLM-L6-v2 embedding model; rankings are nearly identical with other embedding models (Spearman ρ ≥ 0.998).

75 adult volunteers (from 120 recruited, after screening for AI use) did five items of each task; skipped items scored zero.

Scope the authors state

The volunteers are not representative of experts such as scientists, artists or other creators.

Text only; blends use exactly two inputs.

Originality is measured against other answers, so it can’t guarantee an idea is new to all of human knowledge — the benchmark sits between “new to me” and “new to history.”

On this page

Map positions and distances in the Sky, Score and Map sections were computed for this page with all-MiniLM-L6-v2.

The Radio → Milky Way route network, the sample crowd and the whale-shaped-mattress counterexample were written for this page. The van Gogh route, whale groomer drone, market colony, Photosynthesis + Bread descriptions and all reported rates come from the paper.

Figures 4, 7, 8 and the Table 2 illustrations: Schapiro et al., CC BY 4.0 — resized, cropped or redrawn.