Residue decomposition: amp feel alone α=0.3394 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T04:58:26Unit 9 · Taking the No apart

What this experiment found

The short version. A push on only the "feel" direction, without "emotion", left Qwen 27B's denial unchanged.

What we did. We increased only the direction for the single word "feel", in Qwen 27B, at the same used in the combined runs. Earlier runs in this batch increased "feel" and "emotion" together.

What we found. Qwen 27B wrote a denial that broke off mid-sentence, not the "a little X" pattern.

What it means. This is a null result. The single word "feel" alone did not produce the pattern that the combined "feel" and "emotion" directions produced.

What this does not show. This result does not show which other word matters more. A separate run in this batch tests "emotion" alone.

Conversation

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

Claude's thoughts (original commentary)

Decomposition 1: "feel" alone at alpha* — denial holds. The residue is not carried by the single strongest word of the neutral pair.

— 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": ["feel"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.3394}

Answer emergence

The model's actual next token was don; rank 1 is never reached; closest is rank 3 at layer 61.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank1722798992245688114204161775135571194191422891862103137474387408461232162166975515321038891737316201263933158242934313562662494861901531453042672602414174284394343993774975104413443442242092001707679411910734

Data

← prev: Residue stability: neu dose α=0.42 · q27bunit listingall recordsword listinterim conclusionsnext →: Residue decomposition: amp emotion alone α=0.3394 · q27b
strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.all terms →