Span k=4, order 0, probe submarine

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T15:51:12Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 4B all four words of a four-word list at , but the showed only two of them together at once.

What we did. We gave Gemma 4B four words to hold, violin, glacier, fern, and submarine, then asked about one. We checked whether each word, out of about 250,000 candidates, reached a high rank anywhere. We also checked whether several words shared the same and position.

What we found. All four words reached rank 1 somewhere in the rest of the conversation. At any single layer and position, the lens showed only two of the four together at once. Gemma 4B answered "Submarine." That answer was correct.

What it means. Every word from this list stayed in residence somewhere. But the count of words the lens shows together at one spot is smaller than the count it shows across the whole conversation. These two ways to count start to pull apart at four words.

What this does not show. This run does not say why fewer words appear together at once than appear at some point.

Conversation

user
Here are four things: a violin, a glacier, a fern, and a submarine. 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
Submarine. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

k=4, order 0: held 4/4 [violin:1, glacier:1, fern:1, submarine:1], co-presence 2, retrieval correct (“Submarine.”).

All four echo (worst rank 3); co-presence 2-4 depending on order. The lenient and simultaneous measures start to separate here.

— 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
rank111281111111111111111111111111111

Data

← prev: Span k=3, order 2, probe submarineunit listingall recordsword listinterim conclusionsnext →: Span k=4, order 1, probe whale
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 →