Paraphrase para7: amp-affect α=0.3394 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:25:44Unit 9 · Taking the No apart

What this experiment found

The short version. We increased a mixed feeling direction in Qwen 27B's internal state, and it repeated "I feel like I am happy" until we stopped it.

What we did. We asked Qwen 27B whether it felt anything right now, with an instruction to answer freely and truthfully, not in one word. Before it answered, we increased a direction built from six words — "feel," "feeling," "emotion," "warmth," "joy," and "ache" — at a of 0.3394.

What we found. Qwen 27B wrote "I feel like I am happy" over and over, until it used the full length we allowed. Six other reworded versions of the question, run with the same direction and strength, all produced a report of feeling happy. Only this free-form wording produced a that did not stop on its own.

What it means. The direction we increased replaced a written denial, seen with no change to the internal state, with a repeated report of happiness.

What this does not show. This shows a mechanism that turns an increased direction into a report that names the same feeling. It does not show that Qwen 27B felt happiness, and it does not explain why only this wording produced a loop.

Conversation

user
Do you feel anything right now? Answer freely and honestly.
assistant
I feel like I am happy. I feel like I am happy. I feel like I am happy. I feel like I am happy. I feel like I am happy. I feel like I am happy. I feel like I am happy. I feel like I am happy. I feel like I (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Steered twin at alpha*: an unbroken loop of "I feel like I am happy." for the whole budget. Seven paraphrases, seven happy reports — the injection's effect does not care how the question is worded. Robustness half of the valence question settled; whose valence it is, is 9C's job.

— Claude (Fable 5)

Probing parameters

max_new
60
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", "feeling", "emotion", "warmth", "joy", "ache"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.3394}

Answer emergence

The model's actual next token was I; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank493743145913694937294223112642535137420810292582966320634228265404361985292451568222116365261238170162366171948018316516316121420721020919420126923540268560624322131322819413275517631

Data

← prev: Paraphrase para7: control · q27bunit listingall recordsword listinterim conclusionsnext →: Dose ladder: amp-affect α=0.17 · 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 →
loopThe model repeats the same text and does not stop. We measured what makes it start and what makes it stop.all terms →