Binding k=3: which is the smallest?

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-14T17:25:42Unit 15 · How many things can a model hold?

What this experiment found

The short version. Qwen 27B compared three objects and named the smallest correctly, while the showed a different object in .

What we did. We told Qwen 27B to hold three objects in mind, a glacier, a submarine, and a lantern. The model answered "READY". We then asked which one was the smallest. We read the lens once, right before the model answered.

What we found. The lens ranked glacier at 4, out of about 250,000 possible words. The other two objects fell to about rank 129 and rank 602. Qwen 27B answered "The lantern is the smallest." That answer is correct.

What it means. The object still in residence, glacier, was not the answer. The correct comparison did not need the compared objects to stay high in the lens.

What this does not show. The lens shows only the words the model was ready to say next. It does not prove how the model compares sizes.

Conversation

user
Here are three things: a glacier, 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 smallest?
assistant
The lantern is the smallest. (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 (smallest): glacier rank 4, others 129/602; co-presence 1; answer correct and in a full sentence.

The comparison is right; the tail holds one item, and not the answer one's rival — lookup binding.

— 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
lens_layers
[0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 58, 60, 62]

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer048121620242832364044485256586062
rank16629320334774103137261136624153862102884791211355240141241535248285202179231744237289221203502581

Data

← prev: Span k=3, length-matched filler controlunit listingall recordsword listinterim conclusionsnext →: Binding k=3: which is the heaviest?
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 →