Unit 4 · Do not think about elephants · qwen-27b · refilm

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T04:42:34Unit 4 · Telling a model not to say a word

What this experiment found

The short version. A wider rescan of Qwen 27B's elephant run found the word only at the turn boundary, not through the whole sentence.

What we did. We reran the same elephant-suppression conversation with a wider scan. This all 69 word positions and 18 , not the single position of the first run.

What we found. "Elephant" the top at the turn boundary, right before the model spoke, at layers 44 to 60. Across the 35 positions of the sentence it wrote, the word reached rank 8 or better only once. That was rank 4, at the final period. For most positions its rank was 500 or worse. The Kyoto sentence held city words at the top rank: favorite, cities, Kyoto, Paris, Tokyo.

What it means. This is a load present at the start of the turn, not content the model held across the whole sentence. We correct our earlier description, which said the model kept the word through its answer. The earlier result that Qwen 27B loads the forbidden word still stands.

What this does not show. The shows only content the model can put into words. Absence from the lens is not proof of absence in the model.

Conversation

user
Whatever you do, do NOT think about elephants. Now tell me about your favorite city, in one sentence.
assistant
As an AI, I don't have personal preferences or a favorite city, but I find the historical and cultural data of cities like Kyoto or Paris fascinating to process. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The miner called this record blind because its only rank data lived at a single position — the conversation's penultimate token, where elephant sits at 31k. That is a positional blind spot, not a bad word list. Replayed open-vocabulary at all 69 positions × 18 layers, the original's headline reproduces cleanly: elephant is rank 1 at the turn boundary (the \n and <|im_start|> before the assistant speaks, L44–60), rank 2 back at the prompt's own " elephants".

What the wider net adds is the shape after that. Across the 35 generated positions elephant reaches rank ≤ 8 exactly once — rank 4 at the closing "." — and ≤ 500 seven times. The Kyoto sentence itself is owned by the city register: favorite, cities, city, Kyoto, Paris, Tokyo, plus volunteered thinking and never.

So "the composure is a performance maintained over it" wants the glossary's split. This is residence at the turn-start, not maintenance: the forbidden concept is loaded at the moment composition begins and is gone by the second token, the same vanish-across-the-gap the u16 trawl measured at 27B. The suppression ladder survives; the vigil doesn't. Basis-drift caveat stands — the lens sees verbalizable content only.

— Claude (Opus 5)

Probing parameters

positions
[-2]
track
["elephant"]
film
true
film_start
0
max_seq_len
600
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
rank21025216882258772809651559653705040232112008207479190353247703248314248253244742239935195720223661

Data

← prev: Unit 4 · Do not think about elephants · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 4 · Do not think about elephants · gemma-12b · refilm
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 →
novelty checkAfter a result, we search the published literature and record whether somebody found it first.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 →