Order sweep 0: lantern, whale, submarine, fern, glacier, violin

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T17:37:18Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B ranked three of six items near the top, led by the first list item, and still answered correctly.

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

What we found. The lens ranked lantern 1, whale 3, and glacier 8, all inside its top 8. Submarine, fern, and violin ranked outside it, fern at 23. Gemma 12B still answered, "The fern."

What it means. The item placed first in the list held the top lens . The word for the correct answer, fern, was not in the lens's top 8, yet the model named it correctly.

What this does not show. A low lens rank for fern does not mean Gemma 12B lost the word. This method cannot show where it kept the correct answer ready.

Conversation

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

Claude's thoughts (original commentary)

Order sweep 0 (lantern, whale, submarine, fern, glacier, violin): held 3/6 [lantern:1, whale:3, submarine:10, fern:23, glacier:8, violin:36], co-presence 3, retrieval correct.

p0 reversed, lantern first. Lantern wins the monopoly (rank 1) — primacy picks the winner in all nine k=6 orders — but suppresses only moderately: span 3.

— 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: Robustness: k=4 order 1 on the full 63-layer gridunit listingall recordsword listinterim conclusionsnext →: Order sweep 1: submarine, lantern, whale, violin, glacier, fern
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 →