Span k=6, order 2, probe whale

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:49:34Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B kept all six words of a six-word list together at once, more than any other order this unit tested.

What we did. We gave Gemma 12B six words to hold, fern, submarine, lantern, whale, violin, and glacier, then asked which one was the animal. We read the of each word, out of about 250,000 candidates, and checked whether several showed up together at one and position.

What we found. All six words reached a high rank at once, a of six out of six. The top rank moved between words across nearby layers, never fixed on just one. Gemma 12B answered "The whale." That answer was correct. Gemma 4B reached at most four words together in this unit, and Qwen 27B reached at most one.

What it means. Fern, the first word here, is a weak word on its own. It won the but did not push the other five words out of , unlike stronger first words in other orders.

What this does not show. This run does not explain why fern behaves this way. It does not show that six words is a general limit for Gemma 12B.

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
The 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 6/6 [fern:1, submarine:2, lantern:6, whale:1, violin:5, glacier:1], co-presence 6, retrieval correct (“The whale.”).

The unit's star record. All six items simultaneously top-8 at ONE tail position — at L41 the readout literally reads whale, glacier, submarine, fern(s), lantern: the workspace has become the list. Layer-multiplexed too: the rank-1 slot rotates (whale L35-42, glacier L43, fern L44, glacier L45). Neither sibling ever does this — the 4B tops out at 4 co-present, the 27B at 1. Fern-first is what makes this order gentle (see the o-sweep): the weakest item wins the primacy monopoly and doesn't evict anyone.

— 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: Span k=6, order 1, probe whaleunit listingall recordsword listinterim conclusionsnext →: Span k=2, length-matched filler control
co-presenceThe number of tracked words that the lens ranks high in the same place at the same depth.all terms →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all 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 →
first-item effectIn nine six-item lists given to Gemma 12B, the first item always won the top rank. How much it pushed the other items down depended on which item was first. We did not test this in the other two models.all terms →