Order sweep 4: glacier, whale, violin, lantern, fern, submarine

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T17:45:11Unit 15 · How many things can a model hold?

What this experiment found

The short version. With glacier listed first, Gemma 12B again only two of six items near the top of the .

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

What we found. The lens ranked glacier 1 and whale 5, the only two words inside its top 8. Lantern ranked worst, at 253. Gemma 12B still answered "The lantern." Lantern itself was not in the lens's top 8.

What it means. Glacier took the top by being first in the list, then pushed lantern especially far down. A win at the top rank and a strong push down the other words are not the same property.

What this does not show. A very low lens rank for lantern does not mean the model lost it. It named the lantern correctly right after.

Conversation

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

Claude's thoughts (original commentary)

Order sweep 4 (glacier, whale, violin, lantern, fern, submarine): held 2/6 [glacier:1, whale:5, violin:61, lantern:253, fern:27, submarine:17], co-presence 2, retrieval correct.

Fresh order, glacier first. Glacier — the weakest echo in every 4B arm — wins the monopoly here and suppresses hard: violin 61, lantern 253. Span 2. Winning the slot and being crushable are different properties.

— 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 3: violin, submarine, lantern, whale, fern, glacierunit listingall recordsword listinterim conclusionsnext →: Order sweep 5: fern, glacier, violin, submarine, lantern, 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 →