Binding k=3: which is the heaviest?

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:57:44Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B answered a weight comparison correctly while only one of three words stayed strongly in .

What we did. We gave Gemma 12B three words to hold, whale, violin, and fern, then asked which one was the heaviest.

What we found. Whale, the first word in the list, held the top alone, a of one. Violin fell to rank 27 and fern to rank 2. Gemma 12B answered "The whale." That answer was correct.

What it means. This is the again, here in a comparison question rather than a question that only names one item. Whale was both the first word and the correct answer, so a near-empty residence still gave the right reply.

What this does not show. This run does not show whether Gemma 12B can still answer correctly if the heaviest item is not the first word in the list. Other runs in this unit test that case.

Conversation

user

Here are three things: a whale, a violin, 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

Of the things I listed, which one is the heaviest?

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)

Binding k=3 (heaviest): whale monopoly (rank 1), violin evicted to 27, fern 2; co-presence 1; answer correct (whale).

Whale-first collapses here where the same k in part A stayed intact — but the turn-2 text differs, and the int8 instrument is not causal across turn boundaries, so I don't lean on A-vs-B contrasts at this granularity. What's solid: primacy monopoly, correct behavior over a mostly-evicted tail.

— 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	252865	253147	252933	251475	236433	212137	217369	224673	224195	222533	193078	152569	147134	184484	165778	132639	220838	167484	44290	128889	215150	246576	261163	261985	261313	133221	220528	138503	168610	52121	689	112	68	6	29	107	23	19	3	2	2	3	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 →

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 →