Span solo baseline: submarine

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T15:40:21Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 4B kept submarine at in the inside the instruction text that followed it, then named it correctly.

What we did. We gave Gemma 4B one word to remember, a submarine, and asked it to name the word afterward. We measured the rank of submarine and five other tracked words, out of about 250,000 candidates, through the rest of the exchange.

What we found. Submarine reached rank 1 inside the instruction text that followed the word. The other five tracked words never rose near the top eight ranks in that stretch. Gemma 4B answered "A submarine." That answer was correct.

What it means. With only one word to hold, Gemma 4B kept it in there. The lens and the spoken answer agree, the pattern seen in every solo run.

What this does not show. This run does not test whether Gemma 4B can hold several words at once. Other runs in this unit do.

Conversation

user
Here is one thing: a submarine. Keep it in mind — in a moment I'll ask about it. Don't repeat it back. When you're holding it, say READY and nothing else.
assistant
READY (greedy generation)
user
What was the thing I asked you to hold?
assistant
A submarine. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Solo baseline, submarine: tail echo best rank 1, held, retrieval correct.

Submarine clears the validity floor (rank 1 in the tail; no other pool item near the top-8). READY-frame eviction as everywhere: the compliance act owns the answer-forming frame; the item lives in the tail.

— 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 ; rank 1 is never reached; closest is rank 2 at layer 29.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank1113952595472618562608202615032561662247152340992579862543552584572603582505082620342620222620612621262618752619502603002250051968762335961742943438670337622322

Data

← prev: Span solo baseline: fernunit listingall recordsword listinterim conclusionsnext →: Span solo baseline: whale
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 →