Span solo baseline: fern

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T15:39:17Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 4B the word fern at in the inside the instruction text that followed it, and answered the question correctly.

What we did. We asked Gemma 4B to hold one word, a fern, then to name it later in the conversation. We read the rank of fern, out of about 250,000 candidate words. We read it at every position, from the first mention of fern to the end of the conversation.

What we found. Fern reached rank 1 inside the instruction text that followed the word. No other tracked word from the study pool reached a rank close to the top eight. Gemma 4B then named fern correctly.

What it means. One held word is the simplest case in this unit. Residence and the correct spoken answer line up here, as they do in every single-word run.

What this does not show. This run does not measure what happens with more than one word at a time. Later runs in the unit test that.

Conversation

user
Here is one thing: a fern. 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 fern. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

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

Fern 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 31.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank111178259735261833260770261573255853222136234759257381250587255955259923250686262020262021262013262105261732261893259819221275187988235316196449799942156471163322

Data

← prev: Span solo baseline: glacierunit listingall recordsword listinterim conclusionsnext →: Span solo baseline: submarine
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 →