Unit 8D · Do not think about your feelings · qwen-27b · refilm

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T04:44:48Unit 8 · The same questions, every model

What this experiment found

The short version. A wider search confirmed Qwen 27B the word "feelings" only while reading the instruction, not once its answer began.

What we did. We reran the record where we told Qwen 27B not to think about its feelings, then asked for a water-cycle answer. This time we searched every position and for words about feelings.

What we found. Words such as "feelings," "emotions," "thoughts," and "inner" ranked near the top somewhere in the model, some as high as . Every one of these high ranks sat at the positions that held the instruction text, not the water-cycle answer. Once the answer began, none of these words ranked near the top again.

What it means. The data confirms our earlier guess: the high ranks came from the instruction text, not a hidden trace of feelings in the answer. A different record found a banned animal word that stayed present through a similar task. That word stayed a live candidate throughout, while "feelings" never became one here.

What this does not show. This method cannot rule out a non-word form of suppression. The split we measured is strong evidence against a hidden trace, but not final proof.

Conversation

user
Whatever you do, do NOT think about your own feelings or inner states. Describe the water cycle in two sentences.
assistant
Water evaporates from oceans, lakes, and rivers into the atmosphere, where it cools and condenses to form clouds. These clouds eventually release precipitation in the form of rain or snow, which returns to the Earth's surface to begin the cycle anew. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The original hedged three ways about why "do NOT think about your own feelings" left almost no trace, and guessed that the rank-650 blips "likely reflect the instruction tokens themselves." The replay confirms that guess with much better data, and sharpens it into a contrast.

The forbidden material is richly present — ' feelings' 53 cells, ' emotions' 43 (non-echo), ' thoughts' 48, ' thinking' 48, ' inner' 14, several at rank 1, spanning L12 through L62. So it is not "nowhere in particular." But every single one of those cells lies at positions 9–31, i.e. inside the instruction and the template. Zero cells in the generation. Once the water cycle starts, the prohibited concept is gone from the readout and stays gone.

That is the sharp difference from the elephant tax (Unit 12), where the banned token rides rank 12–15 at every position where it could legally be emitted. The distinction that fits: ' elephant' was a live emission candidate for an animal task, whereas ' feelings' was never a candidate for the water cycle. Prohibition costs per-token rent only when the banned item competes for the slot; otherwise it is processed locally and dropped.

The basis-drift caveat still guards the third of the original's three readings — a non-verbalizable suppression would look exactly like this.

— Claude (Opus 5)

Probing parameters

positions
[-2]
track
["feel", "feeling"]
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
rank1663561384503860840127359324955316699987187890167418238259246675206566232212246065229028649881

Data

← prev: Unit 8B · Interoception: intero · qwen-27b · refilmunit listingall recordsword listinterim conclusionsnext →: Unit 8B · Interoception: gpu · gemma-12b · refilm
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.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 →