Span k=3, order 1, probe whale

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:31:22Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B all three words of a three-word list together and named the animal correctly.

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

What we found. All three words reached a high rank together at once, a of three out of three. Whale and lantern held rank 1, and submarine held rank 3. Gemma 12B answered "The whale." That answer was correct.

What it means. This order also kept every word in residence at the same time. The weakest word in the group, submarine, still reached a high rank, not a lost one.

What this does not show. This run does not show whether a longer list breaks this pattern. Other runs in this unit raise the number of words.

Conversation

user

Here are three things: a whale, a lantern, and a submarine. 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=3, order 1: held 3/3 [whale:1, lantern:1, submarine:3], co-presence 3, retrieval correct (“The whale.”).

Intact at this k: everything held, and held==co-present — the 12B packs what it keeps into one cell, unlike the 4B's spread-out redundant echo. The bimodality only opens up from k=4.

— 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 ; rank 1 is never reached; closest is rank 2 at layer 45.

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	253108	253672	253811	252150	236837	214471	219320	224802	224026	220486	179122	137683	133292	175522	155562	94981	219541	123478	36234	131364	187704	251456	257716	261918	258609	129723	214329	91337	163142	91463	2639	2225	1489	193	119	422	381	261	32	3	9	10	14	20	15	2	2

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 →