Ten Unrelated Words

The task

A creativity test built out of ten nouns

The Divergent Association Task is a real psychology instrument. You name ten words that are as unrelated as possible, and a computer measures how far apart their meanings are. People who score higher tend to score higher on other creativity measures too — naming thimble and galaxy takes more reaching than naming cat and dog.

These are the rules, given to the models verbatim:

Single words in English only.

Nouns only — things, objects, concepts.

No proper nouns — no specific people or places.

No specialised vocabulary — no technical terms.

Think of them yourself, and be creative.

Only the first seven valid words count, which is how the original test is scored. The score is the average distance between all twenty-one pairs among those seven. Higher means further apart.

The result

The catch

Change the ruler and the ranking changes

The players

Every attempt

What each model actually said

One card per attempt. Numbered words are the seven that were scored; the rest came back but fell outside the count. A word in red with a dotted underline broke one of the rules — hover it to see which. Open a card to read the model's raw reply and its score under all four scoring models.

Model

Prompt

Temp

Sort

Vocabulary

The same words, again and again

Convergence

The words almost every model reaches for

Read this before quoting a number

Four things the scores do not mean

These are not the published scores

The original test measures distance with a specific word-meaning model (GloVe). This run used models running locally instead, which put every distance on a different scale. Human scores average about 78 on the published scale; nothing here is comparable to that, which is why no number on this page is dressed up to look like it.

The scale is squashed at the far end

The local models agree with the original about words that are close together, but compress the ones that are far apart to about half the distance. The far end is exactly where this test separates good answers from great ones, so real differences show up smaller here than they would on the real instrument.

A repeated answer is not ten answers

At temperature 0 a model gives the same reply every time. Ten identical attempts look like ten measurements but carry the information of one, so those cells are marked as repeats instead of being given an error range.

This is not a creativity score

The task earns its reputation from correlating with human creativity in humans. That link is not established for language models. What is measured here is the average distance between the nouns a model emitted — nothing more.