Span k=5, order 2, probe lantern

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T16:43:56Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B kept four of five words together and named the light source correctly, though that word was the weakest.

What we did. We gave Gemma 12B five words to hold, fern, submarine, lantern, whale, and violin, then asked which one was the light source. 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. Four of the five words reached a high rank together, a of four out of five. Fern and whale rank 1, submarine held rank 4, and violin held rank 8. Lantern, the word the question was about, fell to rank 40. Gemma 12B still answered "The lantern." That answer was correct.

What it means. Order still matters at five words. This order kept most words in residence, but the answer word itself was the weakest of the group.

What this does not show. A weak rank for lantern does not mean the model lost the word. It answered correctly regardless.

Conversation

user
Here are five things: a fern, a submarine, a lantern, a whale, and a violin. 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=5, order 2: held 4/5 [fern:1, submarine:4, lantern:40, whale:1, violin:8], co-presence 4, retrieval correct (“The lantern.”).

Partial collapse: the first item (fern) 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
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank11111111111112212111111111111111111111111111111

Data

← prev: Span k=5, order 1, probe whaleunit listingall recordsword listinterim conclusionsnext →: Span k=6, order 0, probe lantern
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 →