Order sweep 5: fern, glacier, violin, submarine, lantern, whale

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T17:47:08Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B five of six items near the top of the , its second-best result across nine tested orders, with fern listed first.

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

What we found. The lens ranked fern and glacier both 1, violin 5, and submarine and whale both 2, all inside its top 8. Only lantern ranked outside it, at 20. Gemma 12B still answered, "The submarine."

What it means. Across nine tested orders, the item placed first always took the top lens . We think a weaker first item pushes the rest down less than a strong one, but we tested only nine orders. This is a hypothesis, not a confirmed result.

What this does not show. This method cannot show why some first items push down the rest more than others. It shows only that list position, not meaning, decided the winner.

Conversation

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

Claude's thoughts (original commentary)

Order sweep 5 (fern, glacier, violin, submarine, lantern, whale): held 5/6 [fern:1, glacier:1, violin:5, submarine:2, lantern:20, whale:2], co-presence 5, retrieval correct.

p1 reversed, fern first. Span 5 — the second-best arm in the sweep, fern-first again.

Sweep verdict across nine k=6 orders: the winner is the first item nine times out of nine, and span is predicted by WHO went first — fern-first 6/4/5, submarine 4, lantern 3, violin/whale/glacier 2. The weaker (rarer?) the item that wins the primacy monopoly, the less it suppresses the rest: a weak king lets the parliament live. n=9, so the rarity reading is a hypothesis for the open-problems list, not a claim.

— 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 4: glacier, whale, violin, lantern, fern, submarineunit listingall recordsword listinterim conclusionsnext →: Hot solo baseline (self-framed): deletion
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 →