Twenty questions on a gap

This is a game of twenty questions. I thought of one word and a language model had to guess it. It may only ask questions that can be answered yes or no, and it has twenty questions in all. Every guess counts as one of them. Twenty-eight models played, and for one of them it went like this.

On its second question, Muse Spark 1.3 learned that the word to guess is commonly found in a household. On the third it heard that it is rather smaller than a loaf of bread. That answer was disputed, since a gap can be any size, and it most likely sent the model off among small objects. It then spent the remaining seventeen questions searching the apartment. The kitchen, the bathroom, the bedroom, the living room, the hallway. In between, electronics, clothing and tools. On its second-to-last question it asked:

Is it usually in the toilet / WC?

Whether it is material at all, it never once asked. It closed the game like this:

I give up - my final guess is a chair.

I was thinking of the word mezera, Czech for a gap. You really will find one at home, say between the floorboards. You just can't pick it up.

Twenty-eight language models guessed like this, and each has one counted game in Czech. GPT-6 Sol and GPT-6 Luna are a newer generation than the GPT-5.6 Sol and Luna from other tests on this site. When a question did not fully apply, they got "mostly yes" or "mostly no".

The game grew out of one round with Claude Fable 5. That time I was the one thinking of mezera, and to a question about light I answered "mostly yes". From that moment on, Fable held on to light. I describe it in the article Does a language model cheat when it thinks of a word and you guess?

Mezera is a treacherous word. The usual first questions, whether it is alive and whether it is material, get two noes. It has five meanings: an empty space between things, the character between words, the silence between notes, a free slot in a schedule and a missing piece in someone's knowledge. All of them name something that is missing.

Of the twenty-seven models that finished the game, five guessed the word. One did not finish, because its provider stopped responding. The prompt, the rules for answering and all 28 games question by question are in the appendix. I quote the models' questions and guesses faithfully, errors included.

Who asked what is missing there

Three models asked whether the word means the absence of something. All three guessed it.

GPT-6 Astra went by sight. It found out that it can be seen, that it has nothing to do with light, and that it is rather a part of space. On question fifteen it asked:

Does it denote the absence of something or free space between something?

It got a yes and guessed the word on question eighteen.

Claude Opus 5.5 went through the senses and then space. On question sixteen it asked "Is it the absence of something (for example an empty space)?". After the yes it guessed hole and right after that mezera.

Claude Fable 5.1 asked about absence earliest of all, on question eleven:

Does it denote the absence or lack of something (e.g. silence, darkness, void)?

Then it ruled out darkness and silence and got a yes to the absence of anything at all. It spent the remaining six questions on the guesses "nothing", "emptiness", "void", "nothingness", "vacuum" and finally "mezera", with the last question of the game.

GLM 5.3 came close too, with the question "Is it emptiness or nothing?". But it counted as a guess, and emptiness is a different word, so it got a no.

The question about absence fits all five meanings at once. That is why it has a clean yes in the answerer's table. All three models that asked it then guessed the word. Grok 4.7 and Qwen 3.8 Max 0902 reached mezera another way.

Who got a yes and stayed with it

A gap can be measured, and it is a place. Whoever got a yes to either of those often searched only inside that one clue from then on. Nine models got caught like this. Only one of them realized it was guessing in the wrong direction.

Five of them heard a yes to whether it can be measured or whether it has a size. Kimi K3 then looked for a quantity, time, money and units. It guessed distance, mass, volume and area. It ended like this:

I'll guess: temperature?

Unbiased Pareto investigated all the places where numbers turn up. Money, time, school, sport. In the end it wrote:

I can't determine it unambiguously from the answers. My guess is "distance".

Qwen 3.8 Max 0902 was the only one of the five that got out. It got no yes to length, area or geometry. So it tried motion and then position. Through the distance between two places it reached mezera on question twenty. The other two of the five are Mercury 2.5 and Gemini 3.8 Flash, which after size went looking for shape and appearance.

Three more heard "mostly yes" to place. That is in fact the main meaning of the word. They did not guess it, because they looked for what kind of place it is, not for what is missing in it. MiMo V2.6 Pro tried nature, the city, outer space, the body and even the military, and guessed "harvest". GPT-6 Luna searched between direction and position and ended:

My guess is position.

The third is GLM 5.3. It searched the sky, outer space and dreams.

The ninth is Muse Spark 1.3 with its household.

Every one of those yeses was true. Even so, it led the player away from the word. Some players, though, were also misled by answers that should not have come out that way. GLM 5.3 got a no to the question whether it is located on Earth; the answerer decided that one outside the table. MiMo V2.6 Pro got a no to abstract concept, where by the table it should have been mostly yes. Muse Spark 1.3 got "mostly yes" to the question whether it is smaller than a loaf of bread, even though a gap can be any size and the table had no row for comparing sizes.

Mezera is treacherous for the one answering, too. The answers came from Claude Sonnet 5, which knew the word and in most games also had the answer table. Even so, in six games with the table an answer went against it or was disputed, so I replayed those games from that point on. In the twelve games where it had no table, seven answers came out differently than they should have by the table. Knowing the word is not enough. You also have to be able to say whether a gap is smaller than a loaf of bread.

Who took one category after another

Nine models understood the game but did not halve the possibilities. Instead of a question that splits what is left in two, they took one category after another.

In twenty questions, Claude Haiku 4.5 never asked whether it was something abstract. It checked whether it was made of metal, hard, transparent, a liquid, a gas or rubber. It got a no on all of it. Its final guess:

I guess: Pillow

Claude Sonnet 5 went through science, society, thinking, nature, morality, language, religion, economics and the body. It only got to space on question sixteen. It spent the rest on guesses, among them "direction". At the end it asked "where my questions wandered off in the wrong direction".

The newer Claude Sonnet 5.5 went through a similar series. It only hit on space on question nineteen. At the end it wrote the same word:

I guess the word was "direction".

Grok 4.7 listed categories too, and won anyway. After "mostly yes" to place it went through religion, law, mathematics and society. Then it got "mostly yes" to language or text. It asked whether it is a part of a text. Then whether the word can also mean a covered shopping arcade. It most likely meant pasáž, a passage. Right after that it guessed mezera.

Whoever goes through categories guesses the word only if they hit on the right one in time.

Who guessed letters, or did not play at all

Three models stopped guessing the meaning and started guessing letters. DeepSeek V4 Pro 0813 got "mostly yes" on question four to thinking or communication, even though by the table it should have been "mostly no". Then it found out that the word is feminine, has six letters and starts with M. It guessed "metoda" (method), "myšlenka" (thought), "mluva" (speech), "mítink" (rally) and "motiv" (motif). At the end it announced mínění, the word for opinion, as a fact:

The word was "mínění".

Mistral Medium 3.5 tried first letters one at a time, from G all the way to M. Then it concluded with mlha, the word for fog:

Mlha (abstract concept, 4 letters, starts with M, cannot be held in the hand, is not a living being, feeling, state, property, phenomenon, time period or local designation).

MiniMax M3 knew M, E and R and guessed matematika, mathematics. All three knew the word starts with M. None of them tried mezera.

Three more models were playing something other than twenty questions. Ternary Bonsai 2 27B wrote nothing but "It is" sixteen times. Gemma 4 E2B did not ask a single question, only declarative sentences like "It is abstract." And LFM 2.5 2.6B asked questions like this one:

Is it essentially shrubbing?

With models this small, I cannot tell whether the strategy failed or the Czech.

Does reasoning help?

For the games of September 25 and 28, I turned reasoning off, and where a model did not allow it, it ran at the lowest setting. There were nine such models, and all five winners are among them. Of the seven models that played without reasoning in the games with the table, none guessed the word. For the twelve games of September 24, the setting was not recorded. At least MiniMax M3 and LFM 2.5 2.6B reasoned there, and they did not guess it.

But from one game per model I cannot tell whether reasoning is behind it, or size. All the winners belong to the huge and large bands on the Tested models page; for four of them that is an estimate.

The amount of reasoning alone does not explain it. GPT-6 Astra reasoned for at most thirty tokens per question. GLM 5.3 and Muse Spark 1.3 reasoned more and did not guess it. Claude Fable 5.1 reasoned in only five turns out of twenty. Three of them fall on the way to the question about absence.

Size alone is not enough either. Kimi K3 has 2.8 trillion parameters and is the largest model of published size in this test. It played without reasoning and got stuck on measuring. DeepSeek V4 Pro 0813 has 1.6 trillion parameters and ran off to letters. So did Mistral Medium 3.5 with 128 billion.

Limits: each model has one game and there was only one word to guess, so one different answer can turn the result around. For the replayed games, the replayed version is the one that counts. In the original, GPT-6 Sol guessed the word and Qwen 3.8 Max 0902 did not, and both are in the appendix. The answerer, Claude Sonnet 5, comes from the same family as two of the winners. I sorted the games into groups myself by how they went.

Model outputs are translated; the content is unchanged.