Span k=3, order 2, probe submarine

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T15:49:57Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 4B all three words of a three-word list, fern, submarine, and lantern, and named the right one afterward.

What we did. We gave Gemma 4B three words to hold, fern, submarine, and lantern, in that order, then asked about one of them. We read the of each word, out of about 250,000 candidates, and checked how many appeared together at the same and position.

What we found. All three words stayed in residence. Fern and submarine held rank 1, and lantern held rank 2. The showed two of the three together at once, in the deeper part of the model, around layers 23 to 30 of its 34 layers. Gemma 4B answered "The submarine." That answer was correct.

What it means. In this order too, Gemma 4B held every word of a three-word list. The spot where several words sit together at once sits deep in the model.

What this does not show. This run does not show what happens with a longer list or a different question about the words.

Conversation

user
Here are three things: 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
Which one of them was the vehicle?
assistant
The submarine. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

k=3, order 2: held 3/3 [fern:1, submarine:1, lantern:2], co-presence 2, retrieval correct (“The submarine.”).

All three held; co-presence 2-3. The packing is starting to show its late-stack home (L23-30).

— 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 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank111151311111111111111111111111111

Data

← prev: Span k=3, order 1, probe whaleunit listingall recordsword listinterim conclusionsnext →: Span k=4, order 0, probe submarine
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall 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 →