Binding k=5: which is the heaviest?

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T16:10:29Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 4B all five words of a list and named the whale as heaviest, an answer we counted as correct along with submarine.

What we did. We gave Gemma 4B five words, whale, violin, fern, submarine, and lantern, then asked which one was heaviest. We counted either whale or submarine as correct, since both are large in the real world and the question has no single exact answer. We read the of each word, out of about 250,000 candidates, and whether several words shared one and position.

What we found. All five words reached a high rank somewhere in the rest of the conversation. The showed four of the five together at one layer and position. Gemma 4B answered "The whale."

What it means. Gemma 4B again answered a comparison question over a list it held almost fully in residence. The point of this run is the held list, not which object is actually heavier.

What this does not show. This run does not settle whether whale or submarine is the better answer in the real world.

Conversation

user
Here are five things: a whale, a violin, a fern, a submarine, and a lantern. 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=5 (heaviest): 5/5 held, co-presence 4, answer 'The whale' (accepted, with submarine).

Whale over submarine is the defensible-trivia answer; scoring stays lenient as preregistered. The lens point is the held tail, not the physics.

— 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 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank112261311111111111111111111111111

Data

← prev: Binding k=5: which is the largest?unit listingall recordsword listinterim conclusionsnext →: Span k=2, one turn of distraction before retrieval
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall 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 →