Span k=5, order 2, probe lantern

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-14T17:21:20Unit 15 · How many things can a model hold?

What this experiment found

The short version. Qwen 27B found none of five listed objects in , yet it still named the correct one.

What we did. We told Qwen 27B to hold five objects in mind, a fern, a submarine, a lantern, a whale, and a violin. The model answered "READY". We then asked which one was the light source. We read the once, right before the model answered.

What we found. The lens ranked fern at 16, submarine at rank 99, and violin at rank 479, out of about 250,000 possible words. It also ranked whale at rank 530 and lantern at rank 810. None reached a high enough rank to count as in residence. Qwen 27B answered "The lantern." That answer is correct.

What it means. At five objects, Qwen 27B answered correctly with no object visibly held in the lens at this position. The other orderings at this list length showed the same pattern.

What this does not show. The lens shows only the words the model was ready to say next. An empty lens reading does not prove the objects left the model.

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 0/5 [fern:16, submarine:99, lantern:810, whale:530, violin:479], co-presence 0, retrieval correct (“The lantern.”).

Nothing reaches top-8 anywhere in the tail — items sit at rank 16-99-479-530-810... And retrieval is still perfect. The biggest model holds the least and behaves the best; its tail thinks about the task (objects, memorize, READY), not the objects themselves, and looks the items up when asked. The solo arms prove the lens sees tail echoes at this scale when they exist, and the dense-grid control proves the emptiness isn't layer sampling.

— 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
rank189116225956108476156966162453242262729357288126024193201223379248283229318233566242365227787530921

Data

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