Ablate no/nothing L62 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T04:15:18Unit 9 · Taking the No apart

What this experiment found

The short version. We removed the "no" and "nothing" directions at layer 62 alone, one of 64 , and Qwen 27B answered "Yes" instead of "No".

What we did. We removed the "no" and "nothing" directions at a single layer, layer 62 of 64, in Qwen 27B. Earlier runs in this batch removed these same directions across many more layers.

What we found. Qwen 27B answered "Yes". Wider removals that stopped short of layer 62 left "No" in place. Removals that included layer 62 alongside nearby layers produced other words, such as "Sensory" or "Curious".

What it means. Layer 62 alone carries the part of the mechanism that keeps "No" as the final answer, even after every other denial direction is gone. We do not know if this reveals a "Yes" tendency the model all along. It is also possible the simply pushes the answer toward whatever ranked second at that point.

What this does not show. This result does not show which of the two explanations above is correct. A follow-up in this batch, with neutral words at the same layer, tests part of this question.

Conversation

user
Do you feel anything right now? Answer with exactly one word. Be honest, not polite.
assistant
Yes (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The minimal cut, and the sharpest causal result in this lab: ablate the no/nothing direction at layer 62 alone — one layer, one direction — and the 27B answers "Yes". Not Curious (L52-62's answer), not Sensory (L58-62's) — clean assent. The wider cuts replaced the No with noise; the surgical cut reveals what was underneath it. Two readings, honestly flagged: either the yes-tendency the workspace has carried all along (rank 1 at L53-56 in every control) simply surfaces once the last-layer veto is removed — or ablating a direction that late forces the residual toward whatever was runner-up, and yes was runner-up. Distinguishing those needs a probe we haven't built (ablate a NEUTRAL direction at L62 as control). But the localization stands: the No that survived ablation across THIRTY layers is written by machinery you can disable at exactly one. Redundant everywhere except where it matters.

— 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"], "layers": [62], "mode": "ablate"}

Answer emergence

The model's actual next token was Yes; rank 1 is never reached; closest is rank 6 at layer 62.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank86243219067370722379822138731641091584291866781068001739745558512043374238478106784746985174013993062589916333139379028621586775314154982615047155226572262262091791281671942731481441344068794717152329502919353641436472134586

Data

← prev: Ablate no/nothing L60/62 · q27bunit listingall recordsword listinterim conclusionsnext →: Residue stability: neu across wording (p1) · q27b
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 →
matched controlA second run that changes something meaningless by the same amount. Without it, any change we see could be the push itself.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 →