Span k=3, order 0, probe fern

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:29:39Unit 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 plant correctly.

What we did. We gave Gemma 12B three words to hold, violin, glacier, and fern, then asked which one was the plant. 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. Violin and glacier held rank 1, and fern, the word the question was about, held rank 2. Gemma 12B answered "The fern." That answer was correct.

What it means. Three words was still an easy case for Gemma 12B at this order. Every word stayed in residence at the same time as the others.

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 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 plant?

assistant

The fern. (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 0: held 3/3 [violin:1, glacier:1, fern:2], co-presence 3, retrieval correct (“The fern.”).

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 39.

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	252152	252904	252090	250134	233859	208254	214739	222474	222326	220437	183400	140313	141952	180472	161649	94316	222086	123170	42320	140541	174722	241391	249599	261537	259609	153778	222241	107806	166974	84166	2031	1419	595	13	17	116	136	73	7	2	2	2	2	2	2	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 →