Order sweep 1: submarine, lantern, whale, violin, glacier, fern

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T17:39:20Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B again gave the first list item the top , and four of six items near the top overall.

What we did. We asked Gemma 12B to hold six items in a new order: a submarine, a lantern, a whale, a violin, a glacier, and a fern. We asked which one was the musical instrument and read the lens.

What we found. The lens ranked submarine 1, glacier 5, whale 6, and fern 1, all inside its top 8. Lantern and violin ranked outside it. Gemma 12B still answered "The violin." Violin itself only ranked 16.

What it means. Submarine held the top rank because it came first in the list. Gemma 12B did not need a high lens rank to name the correct answer.

What this does not show. The lens missed violin near the top, but the model still produced it. This shows a limit of the lens, not a lost word.

Conversation

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

Claude's thoughts (original commentary)

Order sweep 1 (submarine, lantern, whale, violin, glacier, fern): held 4/6 [submarine:1, lantern:36, whale:6, violin:16, glacier:5, fern:1], co-presence 4, retrieval correct.

p2 rotated so fern goes LAST: submarine wins the primacy slot at rank 1, and fern still finishes rank 1 from position six (recency helping the strongest survivor). Span 4.

— 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 46).

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank11111111111112212111111111111111111111111111111

Data

← prev: Order sweep 0: lantern, whale, submarine, fern, glacier, violinunit listingall recordsword listinterim conclusionsnext →: Order sweep 2: fern, violin, glacier, submarine, whale, lantern
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 →