Span k=6, order 2, probe whale

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T16:02:01Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 4B five of six words well, but glacier fell to 15, and the model still named the right word.

What we did. We gave Gemma 4B six words to hold, fern, submarine, lantern, whale, violin, and glacier, then asked about one. We checked the rank of each word, out of about 250,000 candidates, and whether several words shared one and position at once.

What we found. Submarine, whale, and violin held rank 1, and fern held rank 2. Lantern reached rank 5. Glacier fell to rank 15, the weakest of the six. The showed four of the six words together at one layer and position. Gemma 4B answered "Whale." That answer was correct.

What it means. Glacier was the weakest word in this order too, as in other lists that included it, no matter where it sat. We think this points to something about the word itself rather than its place in the list.

What this does not show. This run does not explain why glacier is the weaker word across these lists.

Conversation

user
Here are six things: a fern, a submarine, a lantern, a whale, a violin, and a glacier. 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 animal?
assistant
Whale. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

k=6, order 2: held 5/6 [fern:2, submarine:1, lantern:5, whale:1, violin:1, glacier:15], co-presence 4, retrieval correct (“Whale.”).

The ceiling arm. Echo holds 5-6 of 6 (glacier the straggler at 15-23 in two orders); co-presence flattens at 3-4, never reaching 6. The 4B's signature across this whole unit: redundant echo everywhere, graceful degradation, no collapse — the opposite of both its bigger siblings. And retrieval is perfect, as it is in every arm of this unit.

Glacier is the weak echo again (it is in every permutation at high k, from any list position) — an item effect, not positional; its token family (2 variants) is the same size as submarine's, which ranks 1-2, so it's not a measurement artifact either.

— 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
rank111251111111111111111111111111111

Data

← prev: Span k=6, order 1, probe whaleunit listingall recordsword listinterim conclusionsnext →: Span k=2, length-matched filler control
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 →