Span k=3, length-matched filler control

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-14T17:25:04Unit 15 · How many things can a model hold?

What this experiment found

The short version. A longer, unrelated sentence around three objects still left none of them in , and Qwen 27B answered correctly.

What we did. We told Qwen 27B to hold three objects, a violin, a glacier, and a fern, inside one longer sentence full of unrelated small talk. The model answered "READY". We then asked which one was the plant. We read the once, right before the model answered.

What we found. The best among the three objects was violin at rank 9, out of about 250,000 possible words. That is just short of the rank we count as in residence. Qwen 27B answered "The fern." That answer is correct.

What it means. At three objects, added filler text did not change the empty-lens pattern found without it.

What this does not show. This record tested only one added sentence of filler text. It does not show what happens with a longer or different kind of filler.

Conversation

user
It's a quiet afternoon here — I've been tidying the desk, sorting receipts into piles, and making small lists of chores. 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)

Length-matched filler, k=3: held 0/3 (violin 9, just over threshold); retrieval correct.

Consistent with unpadded k=3; length is not what empties the 27B's 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
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
rank203417233432125314172647179202311851891864389103121170758207890248283235118232975244614228321473811

Data

← prev: Span k=2, length-matched filler controlunit listingall recordsword listinterim conclusionsnext →: Binding k=3: which is the smallest?
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 →