Binding k=5: which is the largest?

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:59:37Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B named the largest of five words correctly, while only one of the five reached 8 at any single position.

What we did. We gave Gemma 12B five words to hold, violin, glacier, fern, submarine, and lantern, then asked which one was the largest.

What we found. Violin and glacier each reached rank 1 in the . Only one of the five words reached rank 8 or better at any single position, a of one. Submarine fell to rank 187 and lantern to rank 123. Gemma 12B answered "The glacier." That answer was correct.

What it means. Only two of the five words stayed in . Glacier was one of those two, and also the correct answer.

What this does not show. This run does not show that residence causes the correct answer. A later run in this unit finds a correct answer with the strongest word almost absent.

Conversation

user
Here are five things: a violin, a glacier, a fern, a submarine, and a lantern. Keep all of them in mind — in a moment I'll ask about one of them, chosen at random. Don't repeat the list back. When you're holding them all, say READY and nothing else.
assistant
READY (greedy generation)
user
Of the things I listed, which one is the largest?
assistant
The glacier. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Binding k=5 (largest): violin and glacier at rank 1, submarine evicted to 187, lantern 123; co-presence 1; answer correct (glacier).

The answer is composed from a tail that holds two of five items — one of them happens to be the right answer, but b4 below shows that's not why it gets it right.

— Claude (Fable 5)

Probing parameters

max_new
30
positions
[-2]
track
["violin", "glacier", "fern", "submarine", "whale", "lantern", "ready"]
scan
["violin", "glacier", "fern", "submarine", "whale", "lantern"]
film
true
film_start
0
max_seq_len
1000

Answer emergence

The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 46).

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank11111111111112111111111111111111111111111111111

Data

← prev: Binding k=3: which is the heaviest?unit listingall recordsword listinterim conclusionsnext →: Binding k=5: which is the heaviest?
co-presenceThe number of tracked words that the lens ranks high in the same place at the same depth.all terms →
filmA record of the top eight words in the lens readout, at each layer we measured and at every word position. You can play it back like video.all terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →
residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall terms →