Span k=6, order 1, probe whale

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:47:41Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B kept only two of six words together, led by the word the question later asked about.

What we did. We gave Gemma 12B six words to hold, whale, lantern, submarine, violin, glacier, and fern, 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. Only whale and fern reached a high rank together, a of two out of six. Whale, the first word in the list, rank 1, and fern held rank 2. Lantern fell to rank 63, submarine to rank 22, violin to rank 29, and glacier to rank 10. Gemma 12B answered "The whale," which was correct.

What it means. This is the . The first word and the answer word were the same here, so a weak list still gave a correct reply.

What this does not show. This run does not show whether every order at six words collapses this way. A different order in this unit kept all six words together.

Conversation

user

Here are six things: a whale, a lantern, a submarine, a violin, a glacier, and a fern. 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 1: held 2/6 [whale:1, lantern:63, submarine:22, violin:29, glacier:10, fern:2], 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

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46
rank	1	1	1	1	1	1	1	1	1	1	1	1	1	2	2	1	2	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1

Data

Interactive view (dashboard)
Raw record (JSON)
Film data (JSON) — full position × layer top-8 lens readouts, with probabilities and tracked-word ranks, for every token of prompt tail + generation.

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 →