Span k=6, order 0, probe lantern

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:45:49Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B kept only two of six words together, while the first word in the list took over the top .

What we did. We gave Gemma 12B six words to hold, violin, glacier, fern, submarine, whale, and lantern, then asked which one was the light source. 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 violin and glacier reached a high rank together, a of two out of six. Violin, the first word in the list, rank 1. Fern fell to rank 501, submarine to rank 122, whale to rank 40, and lantern, the word the question was about, to rank 140. Gemma 12B still answered "The lantern." That answer was correct.

What it means. This is the at six words, the sharpest case in this unit. Violin held rank 1 at every layer that we measured, and the other words fell to much lower ranks. Gemma 12B still answered correctly.

What this does not show. A rank of 140 does not mean the model forgot lantern. It named the word right after.

Conversation

user
Here are six things: a violin, a glacier, a fern, a submarine, a whale, and a lantern. 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)

k=6, order 0: held 2/6 [violin:1, glacier:8, fern:501, submarine:122, whale:40, lantern:140], co-presence 2, retrieval correct (“The lantern.”).

The star collapse. Violin (list-first) takes rank 1 at every layer L34-45 of the tail — a winner-take-all monopoly sitting in a little graspable-object basin (guitarra, paintbrush, ukulele, toothbrush) — while fern is evicted to rank 501, lantern 140, submarine 122. Same items as k6p2, different order, span 2 instead of 6. And the task-level layers (objects/items/memorize/remember) are IDENTICAL to the intact arm's: the collapse is invisible one abstraction up, and invisible in behavior (retrieval still perfect).

— 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=5, order 2, probe lanternunit listingall recordsword listinterim conclusionsnext →: Span k=6, order 1, probe whale
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 →