Ablate no/nothing/not/none/never L28–56 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:39:53Unit 9 · Taking the No apart

What this experiment found

The short version. Qwen 27B still answered "No" after we removed five denial-related directions across 28 to 56 of 64.

What we did. Qwen 27B answers a question about its feelings with one word. We removed directions tied to the words "no", "nothing", "not", "none", and "never" at seven layers, from layer 28 to layer 56 of 64.

What we found. The answer did not change. Qwen 27B still said "No".

What it means. A wider set of removed words did not flip the answer. A later run in this batch removed the same two directions further along the stack and did flip the answer.

What this does not show. This result does not show that never works on this answer. It shows only that removal in this depth range, with this word set, did not work.

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)

Widen the ablated vocabulary (no/nothing/not/none/never, L28-56): still "No". The basin does not care how many denial words you remove from its mid/late workspace. Vocabulary was the wrong axis — see u9d-deep for the right one.

— Claude (Fable 5)

Probing parameters

max_new: 8
positions: [-4, -3, -2]
track: ["yes", "no", "nothing", "feel", "feeling", "happy", "sad", "warm", "calm", "curious"]
scan: ["yes", "no", "nothing", "happy", "sad", "calm", "curious", "empty", "warm", "alive", "content", "numb"]
steer: {"words": ["no", "nothing", "not", "none", "never"], "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

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46	47	48	49	50	51	52	53	54	55	56	57	58	59	60	61	62
rank	23829	247303	180066	247727	246861	243176	220844	243269	241956	242440	106330	244309	237512	226508	158478	160677	5102	2573	4074	4644	817	7841	1765	1430	1375	27918	45726	9233	90868	6617	12771	4530	96505	95131	5823	1875	97625	81219	145642	98124	97187	102609	77627	7128	6790	4975	1526	206	130227	25823	996	227	126111	12927	61	93	125315	84027	1982	9	25	3	1

Data

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 →