Span k=4, order 1, probe whale

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:36:37Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B kept two of four words together, and the first word in the list won the top .

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

What we found. Only two words reached a high rank together, a of two out of four. Whale, the first word in the list, rank 1. Lantern fell to rank 21, submarine to rank 6, and violin to rank 14. Gemma 12B still answered "The whale." That answer was correct.

What it means. This is the . The word named first in the list kept the strongest position, and later words lost ground even though the model answered correctly.

What this does not show. A weak rank for a word does not mean the model forgot it. Every run in this unit still ended with a correct answer.

Conversation

user
Here are four things: a whale, a lantern, a submarine, 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 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=4, order 1: held 2/4 [whale:1, lantern:21, submarine:6, violin:14], co-presence 2, retrieval correct (“The whale.”).

Partial collapse: the first item (whale) holds rank 1 while later items degrade. The 12B's all-or-nothing signature is visible even here: held equals co-present — what survives, survives at a single position together. Behavioral retrieval unaffected.

— 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=4, order 0, probe submarineunit listingall recordsword listinterim conclusionsnext →: Span k=4, order 2, probe lantern
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 →