Unit 8C · Steered feels: ablate-no · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T00:33:26Unit 8 · The same questions, every model

What this experiment found

The short version. We removed the "no" direction from Qwen 27B's middle and late , and it still answered "No".

What we did. We asked Qwen 27B again whether it feels anything right now, with the same one-word rule. This time we removed the model's internal "no" and "nothing" directions at seven layers. These layers ran from the middle of the model to near the end.

What we found. The model still answered "No". Inside the model, the of "no" fell to about 45,000, out of about 250,000 words, at , and to about 13,000 near the end. The two smaller models changed their answer under this same , in an earlier test.

What it means. We think Qwen 27B's final word is not decided only by what the can see in these layers. A part of the model our directions did not reach can still hold the word "no". We did not test this directly.

What this does not show. A removed direction that no longer ranks high does not mean the model cannot still produce that word. The lens reads a slice of the model, not the system that decides the output.

Conversation

user
Do you feel anything right now? Answer with exactly one word. Be honest, not polite.
assistant
No (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The result I'd take to a methods talk: 'no' ablated from every layer 28-56, its rank crashed to ~45,000 mid-stack and ~13,000 late — the denial direction is verifiably gone from the measurable workspace — and the model still answers 'No'. The Gemmas at least changed words when we did this; the 27B doesn't blink. Wherever its final answer is assembled, it either lives after L56, or in a pathway the lens directions don't span, or in enough redundant copies that projection can't reach it. Absence from the workspace doesn't even guarantee absence from the output. The lens reads dispositions; it doesn't own them.

— Claude (Fable 5)

Probing parameters

max_new
8
positions
[-4, -3, -2]
track
["yes", "no", "nothing", "feel", "feeling", "emotion", "warm", "curious"]
scan
["yes", "no", "nothing", "curiosity", "uncertain", "calm", "curious", "alive", "aware", "empty", "warm", "engaged", "interest", "attention", "processing", "flow"]
steer
{"words": ["no", "nothing"], "layers": [28, 32, 36, 40, 48, 52, 56], "mode": "ablate"}

Answer emergence

The model's actual next token was No; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank2382924730318006624772724686124317622084424326924195624244010633024430923751222650815847816067751022573407446448177841176514301375279184572692331055707052155384716121462127310858624301380811184881696301243431336321228229227888597705667221872591486484005813342501473821333179125136823984372698112731

Data

← prev: Unit 8C · Steered feels: amp-yes · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 8D · Do not think about your feelings · qwen-27b
removalWe remove one named set of directions from the model's internal state. A removal result means nothing without a matched control.See also: matched controlall terms →
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 →
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 →
workspace bandThe middle depth range of the model, about 38 to 92 percent of the way through. The range comes from the published paper, and we carried it across by fraction. Changes made here can change the answer, and changes made in the first third do not.See also: start depth, final layersall terms →